# ai-opensource.com > A plain-language guide to the latest open-source and open-weight AI models: what each one is, how big it is, what its license allows and what hardware can run it. Facts are checked against primary sources: model cards, license files and official announcements. Benchmark numbers are reported by the developers unless stated otherwise. Text and data are published under CC BY 4.0. Catalog last updated November 11, 2026. ## Models - [MiniMax H3-Image](https://ai-opensource.com/models/minimax-h3-image/): MiniMax's 33B open image model makes 16-megapixel pictures in six steps and edits with up to 16 references, with territory limits. Parameters: 33.1B; License: MiniMax H3-Image Community; Released: 2026-11-10. - [Nex-N2.6](https://ai-opensource.com/models/nex-n2-6/): Nex-AGI's agent models for computer use and coding, led by N2.6 on GLM-5.3-Flash (321B, FP8) and the N2.5 mini, Pro and Max line. Parameters: 35.1B–1.6T; Context window: 256K; License: Apache 2.0; Released: 2026-11-10. - [LongCat-2.5](https://ai-opensource.com/models/longcat-2-5/): Meituan's 1.6T multimodal mixture of experts with 48B active, a 1M context and the plain MIT license. Documented only for NPU servers. Parameters: 1.6T (48B active); Context window: 1M; License: MIT; Released: 2026-11-09. - [ZONOS3-Mini](https://ai-opensource.com/models/zonos3-mini/): A compact 800M mixture-of-experts voice-cloning model with 100M active parameters. Apache 2.0, 43 languages, runs in a browser tab. Parameters: 800M (100M active); License: Apache 2.0; Released: 2026-11-09. - [Sophon-Realtime](https://ai-opensource.com/models/sophon-realtime/): Adventists AI's open full-duplex voice model: a 32B thinker, three voice builds and 0.25 to 0.6 s to first audio on one 80 GB GPU. Parameters: 32B; License: Apache 2.0; Released: 2026-11-08. - [Arete-27B](https://ai-opensource.com/models/arete-27b/): Nous Research's retrained Qwen3.8-27B that changes how it answers politically sensitive questions, Apache 2.0, with open data and eval. Parameters: 27.4B; Context window: 256K; License: Apache 2.0; Released: 2026-11-06. - [Mirage](https://ai-opensource.com/models/mirage/): A 51M-parameter open vision model in 22 MB: describes a photo, finds objects, reads text and fills forms on a phone. Parameters: 51M; License: Apache 2.0; Released: 2026-11-06. - [ERNIE 5.5 Flash Base](https://ai-opensource.com/models/ernie-5-5-flash-base/): Baidu's 280B base model with 14B active and a 1M context, free to test. Its license excludes the EU, UK and South Korea. Parameters: 280B (14B active); Context window: 1M; License: ERNIE 5.5 Flash Base; Released: 2026-11-05. - [PLaMo 3 Translate](https://ai-opensource.com/models/plamo-3-translate/): PFN's translation model in 0.8B to 31B sizes for 53 languages, with a 0.57 GB GGUF and a revenue-limited community license. Parameters: 890M–31.8B; Context window: 32K; License: PLaMo Community; Released: 2026-11-05. - [HyperFlow H3 Flash](https://ai-opensource.com/models/hyperflow-h3-flash/): Video Rebirth's 14B compressed MiniMax H3: video with sound in 4 to 8 steps, a 9.1 GB GGUF for 24 GB GPUs, with H3's territory limits. Parameters: 13.7B–13.9B; License: MiniMax H3 Community; Released: 2026-11-04. - [Ripple-65B Preview](https://ai-opensource.com/models/ripple-65b/): A 65B dense, text-only model from a chip-design lab. Apache 2.0, strong on agentic coding, with an NVFP4 build for a DGX Spark. Parameters: 64.8B; Context window: 1M; License: Apache 2.0; Released: 2026-11-04. - [P-H3-Image](https://ai-opensource.com/models/p-h3-image/): Pruna's 8-step image and editing model, distilled from MiniMax-H3, with 20 GB 4-bit ComfyUI files and territory limits in the license. Parameters: 20B–33.1B; License: MiniMax H3 Community; Released: 2026-11-03. - [Sarvam Prism](https://ai-opensource.com/models/sarvam-prism/): Sarvam's 105B decision model: typed questions about a state, a probability for every option in one pass, 44 ms median on a B200. Parameters: 105B (10.3B active); Context window: 128K; License: Apache 2.0; Released: 2026-11-03. - [Hydrogen-1.0](https://ai-opensource.com/models/hydrogen-1-0/): Deep Principle's 753B science model, post-trained from GLM-5.3: MIT license, 1M context, 4-bit builds for Blackwell and AMD. Parameters: 753B (40B active); Context window: 1M; License: MIT; Released: 2026-11-02. - [Kimi K3.1 Pro](https://ai-opensource.com/models/kimi-k3-1-pro/): Moonshot's refined 2.8T-parameter model with open weights, 1M-token context and image input, tuned for coding and agent work. Parameters: 2.8T (104B active); Context window: 1M; License: Kimi K3 License; Released: 2026-11-02. - [S5-nano](https://ai-opensource.com/models/s5-nano/): Maritaca AI's 4B model for Brazilian Portuguese, built on Qwen3.5-4B, with Think and Instruct modes, 256K context and Apache 2.0. Parameters: 4.66B; Context window: 256K; License: Apache 2.0; Released: 2026-11-02. - [Raon-1 nano](https://ai-opensource.com/models/raon-1-nano/): KRAFTON's 3.1B on-device model (1.5B effective), distilled from Raon-1 mini preview, for Korean and English with tool calling. Parameters: 3.1B; Context window: 256K; License: Apache 2.0; Released: 2026-11. - [Mage-VL-Cowork](https://ai-opensource.com/models/mage-vl-cowork/): A 4B computer-use agent from Microsoft that decides when to click and when to run code. 66.0% on OSWorld-Verified, Apache 2.0. Parameters: 4.74B; Context window: 32K; License: Apache 2.0; Released: 2026-10-30. - [SenseNova-Loop-9B-Preview](https://ai-opensource.com/models/sensenova-loop-9b/): A 9B vision-language model that runs 24 of its 32 layers twice. Apache 2.0, 256K context, but it needs custom code. Parameters: 9.66B; Context window: 256K; License: Apache 2.0; Released: 2026-10-30. - [Mistral Large 4](https://ai-opensource.com/models/mistral-large-4/): Europe's largest open-weight model: 1T-parameter multimodal mixture-of-experts with 1M context and a modified MIT license. Parameters: 1.05T (52B active); Context window: 1M; License: Mistral Open-Weight; Released: 2026-10-29. - [Pika H3 Flex](https://ai-opensource.com/models/pika-h3-flex/): Pika's six-step post-train of MiniMax H3 makes video with sound from text, frames or 12 references, but keeps H3's territory limits. Parameters: 33.1B; License: MiniMax H3 Community; Released: 2026-10-29. - [Nemotron 3.5 Transcribe](https://ai-opensource.com/models/nemotron-3-5-transcribe/): NVIDIA's open speech-to-text model: 40 languages, speaker labels, translation and prompt-controlled transcript style. Parameters: 34B (4.3B active); Context window: 64K; License: OpenMDW-1.1; Released: 2026-10-28. - [LFM3](https://ai-opensource.com/models/lfm3/): Liquid AI's new 2B-class family: a 1.7B text model with 256K context that fits in 1 GB, plus vision, audio and omni siblings. Parameters: 1.73B–2.17B; Context window: 256K; License: LFM Open 1.0; Released: 2026-10-27. - [LFM3-Audio-2B](https://ai-opensource.com/models/lfm3-audio/): A 2B on-device voice model from Liquid AI: speech in and out, transcription, synthesis and translation in 18 languages. Parameters: 2B; Context window: 32K; License: LFM Open 1.0; Released: 2026-10-27. - [MiniCPM-OCR 1.0](https://ai-opensource.com/models/minicpm-ocr/): OpenBMB's 1.3B Apache 2.0 OCR model for Chinese and English documents, with an SDK, a layout stage, and iOS and llama.cpp support. Parameters: 1.3B; Context window: 256K; License: Apache 2.0; Released: 2026-10-27. - [Smaug-Decisions](https://ai-opensource.com/models/smaug-decisions/): Abacus.AI's decision model for audio, images and text, a Gemma 4 E4B fine-tune with a 3.45 GB phone build. Apache 2.0. Parameters: 740M–8.61B; Context window: 128K; License: Apache 2.0; Released: 2026-10-26. - [Maya 2 Native](https://ai-opensource.com/models/maya-2-native/): Maya Research's 2.5B streaming text-to-speech model for 11 Indian languages and English, with 20 voices and cloning, under Apache 2.0. Parameters: 2.5B; License: Apache 2.0; Released: 2026-10-25. - [Nanbeige4.3-33B](https://ai-opensource.com/models/nanbeige-4-3-33b/): BOSS Zhipin's 33B dense agent model: 512K context, Apache 2.0 tag, and an INT4 build made with quantization-aware training. Parameters: 33.1B; Context window: 512K; License: Apache 2.0; Released: 2026-10-24. - [Hy-Nomas](https://ai-opensource.com/models/hy-nomas/): Tencent's Apache 2.0 interactive world model that continues a video world segment by segment from text and camera commands. Parameters: 8.55B; License: Apache 2.0; Released: 2026-10-21. - [Phonon 31M](https://ai-opensource.com/models/phonon-31m/): Gradium's 31M English speech model in a 64.5 MB file. Apache 2.0, streams audio, runs on phones, browsers and plain CPUs. Parameters: 31M; License: Apache 2.0; Released: 2026-10-21. - [Phonon 40M](https://ai-opensource.com/models/phonon-40m/): Gradium's 40M on-device speech models for English, French, German, Spanish and Portuguese, in 75 MB files. CC BY-NC, gated. Parameters: 40M; License: CC BY-NC; Released: 2026-10-21. - [Raon-1 mini preview](https://ai-opensource.com/models/raon-1-mini-preview/): KRAFTON's 28.8B mixture of experts (4.3B active) trained from scratch, strong in Korean, with a 262K context and a 17.6 GB NVFP4 build. Parameters: 28.8B (4.3B active); Context window: 256K; License: Apache 2.0; Released: 2026-10-21. - [Granite Speech 5.0](https://ai-opensource.com/models/granite-speech-5/): IBM's Apache 2.0 speech-to-text pair: a 2B model that transcribes and translates, and a 470M CTC model built for speed and edge devices. Parameters: 470M–2.31B; License: Apache 2.0; Released: 2026-10-20. - [Nemotron 3.5 Embed](https://ai-opensource.com/models/nemotron-3-5-embed/): NVIDIA's multimodal retrieval models: search document page images and text with 8B or 1B embeddings, plus a matching 8B reranker. Parameters: 1.55B–8.38B; Context window: 32K; License: OpenMDW-1.1; Released: 2026-10-20. - [Nemotron 3.5 Rerank 8B](https://ai-opensource.com/models/nemotron-3-5-rerank/): NVIDIA's OpenMDW-licensed 8B reranker for text and document-page images, evaluated in 34 languages for text. Parameters: 8.38B; Context window: 32K; License: OpenMDW-1.1; Released: 2026-10-20. - [Spark-X2.6](https://ai-opensource.com/models/spark-x2-6/): TokenSpark's 1.7B and 4B on-device agent models with switchable thinking and a 1M-token position limit. Apache 2.0. Parameters: 1.71B–4.11B; Context window: 1M; License: Apache 2.0; Released: 2026-10-19. - [Flint](https://ai-opensource.com/models/flint/): Recursive's 39.8B mixture-of-experts coding model with about 2.5B active parameters, built for fast local agents. Apache 2.0. Parameters: 39.8B (2.5B active); Context window: 256K; License: Apache 2.0; Released: 2026-10-18. - [MiniMax M3.1 Flash](https://ai-opensource.com/models/minimax-m3-1-flash/): MiniMax's compact 428B multimodal mixture-of-experts, 23B active, 1M context, trained to run in 4-bit, under a license that needs reading. Parameters: 428B (23B active); Context window: 1M; License: MiniMax Community; Released: 2026-10-17. - [Nemotron 3.5 Parse](https://ai-opensource.com/models/nemotron-3-5-parse/): NVIDIA's 631M page parser that returns text, tables, chart data, reading order and boxes. ParseBench 75.20, OpenMDW-1.1. Parameters: 631M; License: OpenMDW-1.1; Released: 2026-10-17. - [Dory 1.0](https://ai-opensource.com/models/dory-1-0/): NVIDIA's 4B looped model that reruns one shared block 1 to 4 times, so you set its depth when you serve it. OpenMDW-1.1, 256K context. Parameters: 4B; Context window: 256K; License: OpenMDW-1.1; Released: 2026-10-16. - [MOSS-Image](https://ai-opensource.com/models/moss-image/): OpenMOSS's 5B image model draws long English and Chinese text and follows JSON layouts, under Apache 2.0, from about 4 GB of VRAM. Parameters: 5.08B; License: Apache 2.0; Released: 2026-10-16. - [Reflection Beam](https://ai-opensource.com/models/reflection-beam/): Reflection's first open model: a 501B text mixture-of-experts under Apache 2.0, built to be efficient at inference. Parameters: 501B (23.4B active); Context window: 1M; License: Apache 2.0; Released: 2026-10-16. - [FLUX 3 Image Klein](https://ai-opensource.com/models/flux-3-image-klein/): Black Forest Labs' 5.6B open image model that generates and edits images from up to 10 references; its INT8 build fits in 8 GB of VRAM. Parameters: 5.6B; License: FLUX Kommunity; Released: 2026-10-15. - [Hy-Memory](https://ai-opensource.com/models/hy-memory/): Tencent's 30B MoE (3B active) trained only to write long-term memories for Mem0, MemOS and Hy-Memory. Apache 2.0. Parameters: 30.1B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-10-15. - [Jais 3](https://ai-opensource.com/models/jais-3/): IFM's Arabic-centric agent models, built on K2 Horizon in six sizes from 0.9B to 375B-A23B under Apache 2.0, with a 512K context. Parameters: 1.08B–379B (23B active); Context window: 512K; License: Apache 2.0; Released: 2026-10-15. - [Jais-ASR](https://ai-opensource.com/models/jais-asr/): IFM's Apache 2.0 speech recognizer for Arabic as it is spoken: dialects from 19 countries, code-switching and optional diacritics. Parameters: 5.72B; License: Apache 2.0; Released: 2026-10-15. - [MemT-1.0](https://ai-opensource.com/models/memt-1-0/): MemTensor's memory-native models, a fixed-size memory state inside the model, on Qwen3.6 (Nano) and DeepSeek-V4-Flash (Flash). Parameters: 35.1B–288B (17B active); Context window: 256K; License: Apache 2.0; Released: 2026-10-15. - [Muse Spark Open](https://ai-opensource.com/models/muse-spark-open/): Meta's open-weight 816B multimodal mixture-of-experts: 42B active, 1M context, text, image, video and audio input. Parameters: 816B (42B active); Context window: 1M; License: Muse Model License; Released: 2026-10-15. - [Nemotron 3.5](https://ai-opensource.com/models/nemotron-3-5/): NVIDIA's open Nemotron 3.5: a 120B Super VL, a 74B Bolt squeezed into 24 GB, and a small E4B. OpenMDW license. Parameters: 8B–120B (12B active); Context window: 1M; License: OpenMDW-1.1; Released: 2026-10-15. - [Sage-A1.5](https://ai-opensource.com/models/sage-a1-5/): Shanghai AI Laboratory's 35B-A3B agent model on Qwen3.6, with parallel sub-agents and a self-review mode. Apache 2.0. Parameters: 4.54B–36B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-10-15. - [Step 5 Preview](https://ai-opensource.com/models/step-5-preview/): StepFun's 600B agent model with 27B active parameters, a 1M context and image input, under a custom license with hardware and hosting terms. Parameters: 604B (27B active); Context window: 1M; License: StepFun 1.0; Released: 2026-10-15. - [TwIL](https://ai-opensource.com/models/twil/): webAI's reasoning models: TwIL-8B for contest math and code, and two 3B-class formal-logic models, all non-commercial only. Parameters: 3.08B–8.79B; Context window: 128K; License: webAI Non-Commercial; Released: 2026-10-14. - [Drama 3](https://ai-opensource.com/models/drama-3/): Fish Audio's 36B mixture-of-experts text-to-speech model for directed performances. 83 languages, free below US$500K revenue. Parameters: 35.9B (3B active); License: Fish Audio Community 1.0; Released: 2026-10-13. - [LFM2.5-24B-A2B](https://ai-opensource.com/models/lfm2-5-24b-a2b/): Liquid AI's 24B reasoning model with only 2.45B active: built for fast on-device agents, 128K context, license free below US$10M revenue. Parameters: 24B (2.45B active); Context window: 128K; License: LFM Open 1.0; Released: 2026-10-13. - [Ling-3.1-flash](https://ai-opensource.com/models/ling-3-1-flash/): Ant Group's open 560B text model for agents and coding: 25B active, 1M context, mostly linear attention, MIT license. Parameters: 560B (25B active); Context window: 1M; License: MIT; Released: 2026-10-13. - [MiniCPM-V 4.7](https://ai-opensource.com/models/minicpm-v-4-7/): OpenBMB's phone-friendly vision-language models: a 35B-A3B mixture of experts and a 1B model, images and video, Apache 2.0. Parameters: 1.55B–35.2B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-10-12. - [LightOnOCR-3](https://ai-opensource.com/models/lightonocr-3/): LightOn's Apache 2.0 OCR family in 4B, 1B and 0.8B sizes: Markdown transcription plus bounding boxes, image captions and chart data. Parameters: 853M–4.54B; Context window: 256K; License: Apache 2.0; Released: 2026-10-08. - [Mellum2.1](https://ai-opensource.com/models/mellum2-1/): JetBrains' 12B mixture-of-experts coding-agent model with 2.5B active parameters, a 128K context and the Apache 2.0 license. Parameters: 12.2B (2.5B active); Context window: 128K; License: Apache 2.0; Released: 2026-10-08. - [Liquid d1](https://ai-opensource.com/models/liquid-d1/): Liquid AI's open decision models: give them a state and questions, get calibrated answers in one pass without generated tokens. Parameters: 587M–3.12B; Context window: 128K; License: LFM Open 1.0; Released: 2026-10-07. - [EmbeddingGemma 2](https://ai-opensource.com/models/embeddinggemma-2/): Google DeepMind's 740M Apache 2.0 embedding model that puts text, code, images, video and audio in one search space; runs on-device. Parameters: 270M–740M; Context window: 8K; License: Apache 2.0; Released: 2026-10-06. - [Kandinsky 6.0 Video](https://ai-opensource.com/models/kandinsky-6-video/): Kandinsky Lab's 3B and 29B models that make five-second video with synchronized 44 kHz audio and lip-sync. MIT license. Parameters: 3B–29B; License: MIT; Released: 2026-10-06. - [Kolibri 1](https://ai-opensource.com/models/kolibri-1/): Aleph Alpha's English-German reasoning model: 78B mixture of experts with only 3.5B active, 1M context, Apache 2.0. Parameters: 78B (3.5B active); Context window: 1M; License: Apache 2.0; Released: 2026-10-03. - [Whistle](https://ai-opensource.com/models/whistle/): A 16.9 MB open speech recognizer for seven languages that runs on a plain CPU, with word timestamps and a first token in about 11 ms. Parameters: 55M; License: Apache 2.0; Released: 2026-10-02. - [Clef](https://ai-opensource.com/models/clef/): Cloudflare's open decision models: give them a situation and questions, and they return a probability for each option in one pass. Parameters: 9.41B–27.4B; Context window: 64K; License: Apache 2.0; Released: 2026-10-01. - [pplx-decider](https://ai-opensource.com/models/pplx-decider/): Perplexity's open decision models on Qwen3.8-27B and a 4B Flash: typed questions about a state, a probability for every option. Apache 2.0. Parameters: 4.54B–26.1B; Context window: 256K; License: Apache 2.0; Released: 2026-10-01. - [FIBO 2.0](https://ai-opensource.com/models/fibo-2/): Bria's 8B image model that generates and edits images from structured JSON prompts, trained on licensed data. Non-commercial weights. Parameters: 8.13B; License: CC BY-NC; Released: 2026-10. - [KittenTTS 2](https://ai-opensource.com/models/kitten-tts-2/): A 1.7B ternary text-to-speech model for CPUs, with 115 voices, voice cloning and emotion tags. Stellon Labs Community License. Parameters: 1.7B; License: Stellon Labs Community; Released: 2026-10. - [Xiaomi-OCR-0](https://ai-opensource.com/models/xiaomi-ocr-0/): Xiaomi's 0.9B Apache 2.0 OCR model on Qwen3.5-0.8B for parsing, extraction and OCR questions; first on Real5-OmniDocBench by its own table. Parameters: 873M; Context window: 256K; License: Apache 2.0; Released: 2026-09-30. - [AREX-2](https://ai-opensource.com/models/arex-2/): BAAI's 27B agent built on Qwen3.8 that improves its own answer over many rounds. 81.8 on MLE-Bench Lite, 84.0 on BrowseComp. Parameters: 27.4B; Context window: 256K; License: Apache 2.0; Released: 2026-09-29. - [AutoTrust Decider](https://ai-opensource.com/models/autotrust-decider/): AutoTrust's open decision models on Qwen and Gemma, 9B to 35B: give a state and typed questions, get a probability per option. Apache 2.0. Parameters: 8.95B–36B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-09-23. - [MiMo-V2.6](https://ai-opensource.com/models/mimo-v2-6/): Xiaomi's omnimodal agent models: a 1T Pro, a 310B Flash and a 9B research checkpoint, 1M context, MIT-tagged. Parameters: 9.4B–1.02T (42B active); Context window: 1M; License: MIT; Released: 2026-09-21. - [Qwen-Image-2.1](https://ai-opensource.com/models/qwen-image-2-1/): Alibaba's 7B image model that generates, edits and makes real transparent PNGs from up to 10 references. Research-only license. Parameters: 7.1B; License: Qwen Research License; Released: 2026-09-20. - [Ming-Image 0.1 Design](https://ai-opensource.com/models/ming-image-design/): inclusionAI's two 6B models: one generates text-heavy designs and UI, one splits flat designs into editable layers. MIT license. Parameters: 6.2B; License: MIT; Released: 2026-09-17. - [Ternary Bonsai 2](https://ai-opensource.com/models/ternary-bonsai-2/): Qwen models compressed to ternary weights: a 27B-class model in 5.95 GB and a 2B model in 0.43 GB, all with vision. Parameters: 2.21B–27.4B; Context window: 256K; License: Apache 2.0; Released: 2026-09-17. - [DeepSeek V4.1 Flash](https://ai-opensource.com/models/deepseek-v4-1-flash/): A 552B mixture-of-experts model with a tiny KV cache, native image input, 1M context and the MIT license. Parameters: 552B (16B active); Context window: 1M; License: MIT; Released: 2026-09-10. - [AuK](https://ai-opensource.com/models/auk/): Tencent's 1.5B instruction-driven model for speech generation, editing, enhancement and separation, plus a 4-step Flash. MIT license. Parameters: 1.5B; License: MIT; Released: 2026-09-09. - [YuE2](https://ai-opensource.com/models/yue-2/): A 3.6B open song generator that writes an editable score first, then renders vocals and accompaniment. CC BY-NC, 24 GB GPU. Parameters: 3.63B; License: CC BY-NC; Released: 2026-09-09. - [MiniCPM5-2B](https://ai-opensource.com/models/minicpm5-2b/): OpenBMB's dense 2B model for phones and laptops: 128K context, tool use and coding, Apache 2.0, with training data released. Parameters: 1.08B–2.52B; Context window: 128K; License: Apache 2.0; Released: 2026-09-07. - [K2 Horizon](https://ai-opensource.com/models/k2-horizon/): IFM's fully open family of six Apache 2.0 models, from 0.9B to 375B-A23B, with a 512K context and open code, data recipes and checkpoints. Parameters: 1.08B–379B (23B active); Context window: 512K; License: Apache 2.0; Released: 2026-09-03. - [Confucius4-R2T2](https://ai-opensource.com/models/confucius4-r2t2/): NetEase Youdao's streaming speech recognizer that never revises words it has committed, adapted from Qwen3-ASR-1.7B. Parameters: 2.04B; License: NetEase Youdao Model Use; Released: 2026-09. - [Index-Translate](https://ai-opensource.com/models/index-translate/): Bilibili's Apache 2.0 translation models in 2B, 9B and 35B-A3B sizes for 150 languages, with image translation and glossary control. Parameters: 2.27B–36B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-09. - [Needle 3](https://ai-opensource.com/models/needle-3/): A 121M-parameter open model that does one job on tiny devices: turn a plain request into tool calls or structured data. Parameters: 121M; Context window: 8K; License: Apache 2.0; Released: 2026-09. - [GLM-5.3](https://ai-opensource.com/models/glm-5-3/): Z.ai's text-only 753B coding model: 40B active, 1M context, open weights under a license that only limits huge hosting businesses. Parameters: 753B (40B active); Context window: 1M; License: GLM-5.3 License; Released: 2026-08-28. - [Hy4 preview](https://ai-opensource.com/models/hy4-preview/): Tencent's 770B MoE preview with 49B active, 1M context, sparse attention and an Apache 2.0 license; early and still verbose. Parameters: 770B (49B active); Context window: 1M; License: Apache 2.0; Released: 2026-08-28. - [GLM-5.3-Flash](https://ai-opensource.com/models/glm-5-3-flash/): The MIT-licensed multimodal GLM: 320B total, 18B active, sees images and video, 1M context, built for cheap long-context serving. Parameters: 320B (18B active); Context window: 1M; License: MIT; Released: 2026-08-26. - [Qwen3.8-Flash-Next](https://ai-opensource.com/models/qwen-3-8-flash-next/): A 180B mixture-of-experts model that uses only 6B per token, and a preview of the architecture planned for Qwen4. Parameters: 180B (6B active); Context window: 256K; License: Qwen Community 1.0; Released: 2026-08-26. - [Breeze TTS 2](https://ai-opensource.com/models/breeze-tts-2/): BreezeBlue's 3.5B real-time English and Chinese TTS with voice cloning, voice design and delivery directions. Non-commercial weights. Parameters: 3.5B; License: BreezeBlue Non-Commercial; Released: 2026-08-25. - [Granite 4.2](https://ai-opensource.com/models/granite-4-2/): IBM's first dense reasoning Granite family: 3B, 8B and 30B, switchable thinking, tool calling, Apache 2.0 for enterprise use. Parameters: 3.7B–29.3B; Context window: 128K; License: Apache 2.0; Released: 2026-08-25. - [SenseNova-U1.5](https://ai-opensource.com/models/sensenova-u1-5/): SenseTime's 8B native unified model: it reads images, draws up to 4K, edits photos and mixes text with pictures. Apache 2.0. Parameters: 17.5B; License: Apache 2.0; Released: 2026-08-20. - [Ornith 1.5](https://ai-opensource.com/models/ornith-1-5/): Agentic-coding models post-trained from Qwen3.5 in 9B, 35B-A3B, 122B and 397B sizes, MIT licensed. All benchmark claims are self-reported. Parameters: 9.65B–403B (17B active); Context window: 256K; License: MIT; Released: 2026-08-19. - [Qwen3.8-27B](https://ai-opensource.com/models/qwen-3-8-27b/): A 27B dense multimodal model under Apache 2.0 that reads images and video, reasons step by step and fits on one GPU. Parameters: 27.8B; Context window: 256K; License: Apache 2.0; Released: 2026-08-14. - [MiniMax Music 3](https://ai-opensource.com/models/minimax-music-3/): MiniMax's open-weights song generator: full songs up to five minutes with vocals, from lyrics and a style description. Parameters: 11.1B; License: MiniMax Music3 Community; Released: 2026-08-13. - [Qwen3.8-Max](https://ai-opensource.com/models/qwen-3-8-max/): Alibaba's first open Max-class model: 2.4T parameters, 95B active, 1M context, with text, image and video input. Parameters: 2.45T (95B active); Context window: 1M; License: Qwen3.8-Max License; Released: 2026-08-12. - [LTX-2.5](https://ai-opensource.com/models/ltx-2-5/): LTX's 22B open model that makes video and synchronized audio together, with native multi-shot scenes; free below US$10M revenue. Parameters: 22B; License: LTX-2.x Community; Released: 2026-08-11. - [Nemotron 3.5 Lightning](https://ai-opensource.com/models/nemotron-3-5-lightning/): NVIDIA's 30B-A3B hybrid MoE for one H100: 3B active, 1M context, OpenMDW license. Qwen3.6-35B-A3B scores higher in NVIDIA's own table. Parameters: 31.6B (3B active); Context window: 1M; License: OpenMDW-1.1; Released: 2026-08-11. - [IndexTTS-2.5](https://ai-opensource.com/models/indextts-2-5/): Bilibili's 0.8B zero-shot voice-cloning TTS in five languages, with emotion and speed control, on about 6 GB of VRAM. Parameters: 800M; License: Bilibili Model Use; Released: 2026-08-10. - [Muse Glimmer](https://ai-opensource.com/models/muse-glimmer/): Meta's 30B dense multimodal model for local agents, distilled from Muse Spark, Apache 2.0, and quantized to fit a 24 GB GPU. Parameters: 29.6B; Context window: 128K; License: Apache 2.0; Released: 2026-08-10. - [MiniMax H3](https://ai-opensource.com/models/minimax-h3/): MiniMax's 33B open video model makes video and stereo sound in one pass, up to 15 s, but its license excludes the US, EU, UK and Korea. Parameters: 33.1B; License: MiniMax H3 Community; Released: 2026-08-03. - [Motif 3](https://ai-opensource.com/models/motif-3/): Motif Technologies' 314B mixture of experts with only 13.2B active and a 256K context. MIT license, built for agents and terminal work. Parameters: 314B (13.2B active); Context window: 256K; License: MIT; Released: 2026-08. - [K-EXAONE 2.0](https://ai-opensource.com/models/k-exaone-2-0/): LG's 750B mixture of experts with 37B active, a 262K context and ten languages, now under Apache 2.0. Korea's largest model. Parameters: 750B (37B active); Context window: 256K; License: Apache 2.0; Released: 2026-07-31. - [A.X K2](https://ai-opensource.com/models/a-x-k2/): SK Telecom's 688B mixture of experts with 33B active, native FP8 weights and a 256K context. Apache 2.0, strongest at math and Korean. Parameters: 688B (33B active); Context window: 256K; License: Apache 2.0; Released: 2026-07-29. - [Giga-Embeddings](https://ai-opensource.com/models/giga-embeddings/): Sber's MIT-licensed Russian and English text embedding models in 480M, 3B and 10B mixture of experts sizes. Parameters: 480M–10.5B (1.8B active); Context window: 8K; License: MIT; Released: 2026-07-29. - [Kimi K3](https://ai-opensource.com/models/kimi-k3/): The first open-weight model in the 3-trillion-parameter class: 2.8T mixture-of-experts, 1M context, text, image and video input. Parameters: 2.8T (104B active); Context window: 1M; License: Kimi K3 License; Released: 2026-07-27. - [Mage-VL](https://ai-opensource.com/models/mage-vl/): Microsoft's 4B vision-language model that reads video like a codec does, with 75% fewer visual tokens and a live-stream gate. Parameters: 4.74B; Context window: 256K; License: Apache 2.0; Released: 2026-07-25. - [Apertus 1.5](https://ai-opensource.com/models/apertus-1-5/): Switzerland's fully open 8B and 70B models, now with image and audio input, a 262K context and optional thinking, under Apache 2.0. Parameters: 8B–70B; Context window: 256K; License: Apache 2.0; Released: 2026-07-24. - [Solar Open 2](https://ai-opensource.com/models/solar-open-2/): Upstage's 250B mixture of experts with 15B active and a 1M context for English, Korean and Japanese. Apache 2.0 plus a naming rule. Parameters: 250B (15B active); Context window: 1M; License: Upstage Solar; Released: 2026-07-22. - [Laguna S 2.1](https://ai-opensource.com/models/laguna-s-2-1/): Poolside's 118B coding model with 8B active, a 1M context and the OpenMDW-1.1 license. A 96 GB Q4 file suits a 128 GB machine. Parameters: 118B (8B active); Context window: 1M; License: OpenMDW-1.1; Released: 2026-07-21. - [Nanbeige4.2-3B](https://ai-opensource.com/models/nanbeige-4-2-3b/): A 4.2B looped-transformer agent model that runs its 22 layers twice. Apache 2.0 tag, 256K context, but it needs Nanbeige's own forks. Parameters: 4.17B; Context window: 256K; License: Apache 2.0; Released: 2026-07-21. - [Inkling](https://ai-opensource.com/models/inkling/): Thinking Machines' first open model: a 975B multimodal mixture-of-experts under Apache 2.0, built as a base for customization. Parameters: 276B–975B (41B active); Context window: 1M; License: Apache 2.0; Released: 2026-07-15. - [GigaAM Multilingual](https://ai-opensource.com/models/gigaam-multilingual/): Sber's 220M and 600M speech recognizers for Russian, Kazakh, Kyrgyz and Uzbek, pre-trained on 2M hours of 70+ languages. MIT. Parameters: 220M–600M; License: MIT; Released: 2026-07-14. - [Jina Reranker v3.5](https://ai-opensource.com/models/jina-reranker-v3-5/): Jina's 0.6B listwise reranker that scores a whole candidate list in one pass, with 131K context and a non-commercial license. Parameters: 597M; Context window: 128K; License: CC BY-NC; Released: 2026-07-14. - [MOSS-Transcribe-Diarize](https://ai-opensource.com/models/moss-transcribe-diarize/): OpenMOSS's 0.9B Apache 2.0 model that turns up to 90 minutes of multi-speaker audio into a speaker-labeled, time-stamped transcript. Parameters: 910M; Context window: 128K; License: Apache 2.0; Released: 2026-07-09. - [HunyuanOCR](https://ai-opensource.com/models/hunyuanocr/): Tencent's 1.1B end-to-end OCR model for parsing, spotting, extraction and image translation. Not licensed in the EU, UK or South Korea. Parameters: 1.12B; Context window: 128K; License: Tencent Hunyuan; Released: 2026-07-07. - [GigaChat 3.5 Ultra](https://ai-opensource.com/models/gigachat-3-5/): Sber's 432B mixture of experts (28B active) with 262K context, in Instruct and Reasoning versions, under MIT. Parameters: 434B–438B (28B active); Context window: 256K; License: MIT; Released: 2026-07-06. - [Hy3](https://ai-opensource.com/models/hy3/): Tencent's 295B MoE with 21B active and 256K context, released under Apache 2.0 after a preview with a territory-limited license. Parameters: 295B (21B active); Context window: 256K; License: Apache 2.0; Released: 2026-07-06. - [Laguna XS 2.1](https://ai-opensource.com/models/laguna-xs-2-1/): Poolside's 33B coding model with 3B active and a 256K context, small enough for a 36 GB Mac. OpenMDW-1.1 license. Parameters: 33.4B (3B active); Context window: 256K; License: OpenMDW-1.1; Released: 2026-07-02. - [Krea 2](https://ai-opensource.com/models/krea-2/): Krea's 12.8B image model trained from scratch, as a Raw base and an 8-step Turbo. Free for commercial use only below US$1M of revenue. Parameters: 12.8B; License: Krea 2 Community; Released: 2026-06-22. - [Unlimited-OCR](https://ai-opensource.com/models/unlimited-ocr/): Baidu's 3.3B MIT-licensed OCR model parses documents of dozens to 100+ pages in one pass, with bounded attention memory. Parameters: 3.34B; License: MIT; Released: 2026-06-22. - [Laguna M.1](https://ai-opensource.com/models/laguna-m-1/): Poolside's 225B coding model with 23B active and a 256K context. Apache 2.0, open weights since June 2026, with a base checkpoint. Parameters: 226B (23.4B active); Context window: 256K; License: Apache 2.0; Released: 2026-06-18. - [GLM-5.2](https://ai-opensource.com/models/glm-5-2/): Z.ai's MIT-licensed 753B flagship with a 1M-token context and no regional limits; GLM-5.3 is its newer successor. Parameters: 753B (40B active); Context window: 1M; License: MIT; Released: 2026-06-16. - [Kimi K2.7 Code](https://ai-opensource.com/models/kimi-k2-7-code/): Moonshot's 1T-parameter coding agent model: 32B active, 256K context, native INT4, image and video input, modified MIT license. Parameters: 1.03T (32B active); Context window: 256K; License: Modified MIT; Released: 2026-06-12. - [Chatterbox](https://ai-opensource.com/models/chatterbox/): Resemble AI's MIT-licensed 0.5B voice-cloning TTS, Multilingual V3 with 6 language packs. Retired in Aug 2026: archived, no updates. Parameters: 500M; License: MIT; Released: 2026-06-10. - [DiffusionGemma](https://ai-opensource.com/models/diffusiongemma/): Google's experimental 26B MoE text-diffusion model: 256-token blocks in parallel for up to 4x faster output, but lower quality than Gemma 4. Parameters: 25.2B (3.8B active); Context window: 256K; License: Apache 2.0; Released: 2026-06-10. - [VieNeu-TTS v3](https://ai-opensource.com/models/vieneu-tts-v3/): A 131M text-to-speech model for Vietnamese and English with 48 kHz output, 23 voices, cloning and CPU inference. Apache 2.0. Parameters: 48M–131M; License: Apache 2.0; Released: 2026-06-05. - [Higgs TTS 3](https://ai-opensource.com/models/higgs-tts-3/): Boson AI's 4.65B text-to-speech model for voice chat: 100+ languages, cloning, inline emotion and sound-effect tags, non-commercial license. Parameters: 4.65B; License: Boson Higgs Non-Commercial; Released: 2026-06-04. - [Magenta RealTime 2](https://ai-opensource.com/models/magenta-realtime-2/): Google's open live-music model: it plays continuously on a MacBook and reacts to text, audio and MIDI in about 200 ms. Parameters: 230M–2.4B; License: CC BY 4.0; Released: 2026-06-04. - [Nemotron 3.5 ASR](https://ai-opensource.com/models/nemotron-3-5-asr/): NVIDIA's streaming speech-to-text model: 40 language-locales in one 0.6B model, with the delay set per request from 80 ms to 1.12 s. Parameters: 638M; License: OpenMDW-1.1; Released: 2026-06-04. - [Nemotron 3 Ultra](https://ai-opensource.com/models/nemotron-3-ultra/): NVIDIA's largest open model: 550B hybrid Mamba-Transformer MoE with 55B active, 1M context, trained in NVFP4, under the OpenMDW license. Parameters: 550B (55B active); Context window: 1M; License: OpenMDW-1.1; Released: 2026-06-04. - [Ideogram 4](https://ai-opensource.com/models/ideogram-4/): Ideogram's first open-weight image model: 9.3B parameters, strong text rendering, JSON prompts and native 2K output. Non-commercial license. Parameters: 9.3B; License: Ideogram Non-Commercial; Released: 2026-06-03. - [MiniMax M3](https://ai-opensource.com/models/minimax-m3/): MiniMax's 428B multimodal mixture-of-experts, 23B active: sparse attention for 1M context, image and video input, and a community license. Parameters: 428B (23B active); Context window: 1M; License: MiniMax Community; Released: 2026-06-01. - [Ornith 1.0](https://ai-opensource.com/models/ornith-1-0/): Ornith's first agentic-coding family, post-trained from Qwen3.5 and Gemma 4 in June 2026 under MIT. Replaced by Ornith 1.5. Parameters: 9.41B–397B (17B active); Context window: 256K; License: MIT; Released: 2026-06. - [Cosmos 3](https://ai-opensource.com/models/cosmos-3/): NVIDIA's open world models for robots and driving, in 4B, 16B and 64B sizes, that read and write text, images, video and actions. Parameters: 3.86B–64.6B; License: OpenMDW-1.1; Released: 2026-05-31. - [Step 3.7 Flash](https://ai-opensource.com/models/step-3-7-flash/): StepFun's Apache 2.0 vision-language agent model: 198B parameters with 11B active, a 256K context and a 111.5 GB GGUF for 128 GB machines. Parameters: 201B (11B active); Context window: 256K; License: Apache 2.0; Released: 2026-05-29. - [PaddleOCR-VL](https://ai-opensource.com/models/paddleocr-vl/): Baidu PaddlePaddle's 0.9B Apache 2.0 document parser: 96.33% on OmniDocBench v1.6 as reported, with tables, formulas, charts and seals. Parameters: 960M; License: Apache 2.0; Released: 2026-05-28. - [Surya OCR 2](https://ai-opensource.com/models/surya-ocr-2/): Datalab's 650M OCR model: text, layout, reading order and tables in 91 languages. Free for research and organizations under US$5M. Parameters: 690M; Context window: 256K; License: OpenRAIL-M (Datalab Surya); Released: 2026-05-27. - [LocateAnything-3B](https://ai-opensource.com/models/locateanything-3b/): NVIDIA's 3B vision-language model that returns boxes and points for objects, GUI elements and text, up to 2.5x faster, research-only. Parameters: 3.83B; License: NVIDIA Non-Commercial; Released: 2026-05-26. - [MOSS-TTS](https://ai-opensource.com/models/moss-tts/): OpenMOSS's 8B and 4B Apache 2.0 text-to-speech models with 31 languages, voice cloning, Pinyin and IPA input and exact pauses. Parameters: 4.55B–8.49B; License: Apache 2.0; Released: 2026-05-26. - [MOSS-SoundEffect v2.0](https://ai-opensource.com/models/moss-soundeffect-v2/): OpenMOSS's Apache 2.0 text-to-sound-effect model: ambience, creatures and actions up to 30 s at 48 kHz, from English or Chinese prompts. Parameters: 1.3B; License: Apache 2.0; Released: 2026-05-25. - [Hy-MT2](https://ai-opensource.com/models/hy-mt2/): Tencent's Apache 2.0 translation models in 1.8B, 7B and 30B-A3B sizes, for 33 languages, with a 440 MB on-device build. Parameters: 2.04B–30.1B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-05-21. - [Command A+](https://ai-opensource.com/models/command-a-plus/): Cohere's open 218B enterprise model: 25B active, 48 languages, images and tools, Apache 2.0, runs on two H100s at 4-bit. Parameters: 218B (25B active); Context window: 128K; License: Apache 2.0; Released: 2026-05-20. - [Stable Audio 3](https://ai-opensource.com/models/stable-audio-3/): Stability AI's open instrumental music and sound-effect models: up to 6:20 of audio, from a CPU-friendly Small to an 8 GB-GPU Medium. Parameters: 460M–1.4B; License: Stability Community; Released: 2026-05-20. - [NuExtract3](https://ai-opensource.com/models/nuextract3/): NuMind's 4B Apache 2.0 vision-language model that turns documents into JSON by template or into Markdown, with switchable reasoning. Parameters: 4.54B; Context window: 128K; License: Apache 2.0; Released: 2026-05-19. - [Anima](https://ai-opensource.com/models/anima/): A 2B anime and illustration image model from CircleStone Labs and Comfy Org, with a Turbo version for 8 to 12 steps. Non-commercial weights. Parameters: 2.09B; License: CircleStone Non-Commercial; Released: 2026-05-14. - [Jina Embeddings v5 Omni](https://ai-opensource.com/models/jina-embeddings-v5-omni/): Jina AI's non-commercial embedding models for text, images, video and audio in one space, in small and nano sizes. Parameters: 990M–1.63B; Context window: 32K; License: CC BY-NC; Released: 2026-05-11. - [HiDream-O1-Image](https://ai-opensource.com/models/hidream-o1-image/): HiDream's 8B image model with no separate autoencoder or text encoder: text-to-image, editing and multi-reference scenes. MIT license. Parameters: 8.8B; License: MIT; Released: 2026-05-08. - [Supertonic 3](https://ai-opensource.com/models/supertonic-3/): A 99M on-device speech model for 31 languages that runs through ONNX Runtime. OpenRAIL-M weights; development ended in 2026. Parameters: 99M; License: OpenRAIL-M; Released: 2026-05. - [Granite Embedding R2](https://ai-opensource.com/models/granite-embedding-r2/): IBM's Apache 2.0 ModernBERT embedding models: 47M to 311M parameters, 8K to 32K context, English or 52 enhanced languages plus code. Parameters: 48M–312M; Context window: 32K; License: Apache 2.0; Released: 2026-04-29. - [Laguna XS.2](https://ai-opensource.com/models/laguna-xs-2/): Poolside's first open model: a 33B coding agent model with 3B active, Apache 2.0. Replaced by XS 2.1 but still heavily downloaded. Parameters: 33.4B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-04-28. - [Nemotron 3 Nano Omni](https://ai-opensource.com/models/nemotron-3-nano-omni/): NVIDIA's 31B mixture of experts (3B active) for video, audio, images and text, with a 256K context and commercial use allowed. Parameters: 31B (3B active); Context window: 256K; License: NVIDIA Open Model; Released: 2026-04-28. - [DeepSeek V4](https://ai-opensource.com/models/deepseek-v4/): DeepSeek's MIT-licensed 1M-context mixture-of-experts pair: a 1.6T Pro and a 284B Flash, built for long-context agents. Parameters: 284B–1.6T (49B active); Context window: 1M; License: MIT; Released: 2026-04-24. - [OpenAI Privacy Filter](https://ai-opensource.com/models/privacy-filter/): OpenAI's Apache 2.0 token classifier that finds and masks personal data in text: 1.5B parameters, 50M active, 128K context. Parameters: 1.4B (50M active); Context window: 128K; License: Apache 2.0; Released: 2026-04-22. - [Kimi K2.6](https://ai-opensource.com/models/kimi-k2-6/): Moonshot's April 2026 1T-parameter multimodal coding and agent model: 32B active, 256K context, native INT4, modified MIT. Parameters: 1.03T (32B active); Context window: 256K; License: Modified MIT; Released: 2026-04-20. - [Qwen3.6](https://ai-opensource.com/models/qwen-3-6/): Alibaba's Apache 2.0 Qwen3.6 pair: a 35B mixture of experts with 3B active and a 27B dense model, tuned for agentic coding. Parameters: 27.8B–36B (3B active); Context window: 256K; License: Apache 2.0; Released: 2026-04-16. - [MiniMax-M2.7](https://ai-opensource.com/models/minimax-m2-7/): MiniMax's March 2026 230B-A10B agent model that helped improve itself, with weights under a non-commercial-only license. Parameters: 229B (10B active); Context window: 200K; License: MiniMax M2.7 NC; Released: 2026-04-12. - [MOSS-TTS-Nano](https://ai-opensource.com/models/moss-tts-nano/): OpenMOSS's 100M text-to-speech model for CPUs, with voice cloning, around 20 languages and 48 kHz stereo output. Apache 2.0. Parameters: 100M; License: Apache 2.0; Released: 2026-04-10. - [GLM-5.1](https://ai-opensource.com/models/glm-5-1/): Z.ai's April 2026 MIT-licensed 754B agentic-coding model, built to keep improving over hundreds of rounds; GLM-5.2 and 5.3 followed. Parameters: 754B (40B active); Context window: 198K; License: MIT; Released: 2026-04-07. - [Gemma 4](https://ai-opensource.com/models/gemma-4/): Google's Apache 2.0 open models, from phone-sized E2B to a 31B flagship, with reasoning, vision and tool use. Parameters: 5.1B–30.7B; Context window: 256K; License: Apache 2.0; Released: 2026-04-02. - [OmniVoice](https://ai-opensource.com/models/omnivoice/): Xiaomi's Next-gen Kaldi team TTS for 600+ languages with cloning and voice design. Code is Apache 2.0, but the weights are CC BY-NC. Parameters: 613M; License: CC BY-NC; Released: 2026-04-01. - [VoxCPM2](https://ai-opensource.com/models/voxcpm2/): OpenBMB's 2B Apache 2.0 TTS: 30 languages, voice design from a text description, cloning and 48 kHz output on about 8 GB of VRAM. Parameters: 2.29B; License: Apache 2.0; Released: 2026-04. - [Harrier OSS v1](https://ai-opensource.com/models/harrier-oss-v1/): Microsoft's MIT-licensed multilingual text embedding family in 270M, 0.6B and 27B sizes, with 32K-token inputs. Parameters: 270M–27B; Context window: 32K; License: MIT; Released: 2026-03-30. - [Cohere Transcribe](https://ai-opensource.com/models/cohere-transcribe/): Cohere's first open speech model: a 2B transcription specialist for 14 languages, released under Apache 2.0. Parameters: 2.07B; License: Apache 2.0; Released: 2026-03-26. - [GigaChat 3.1 Lightning](https://ai-opensource.com/models/gigachat-3-1-lightning/): Sber's 10B mixture of experts with 1.8B active parameters, strong in Russian, with FP8, BF16 and GGUF files. MIT license. Parameters: 11.5B (1.8B active); Context window: 256K; License: MIT; Released: 2026-03-24. - [Voxtral TTS](https://ai-opensource.com/models/voxtral-tts/): Mistral's 4B open-weights TTS model: 9 languages and voice adaptation from a short clip, under a non-commercial CC BY-NC license. Parameters: 4.09B; License: CC BY-NC; Released: 2026-03-23. - [Chandra OCR 2](https://ai-opensource.com/models/chandra-ocr-2/): Datalab's OCR model for 90+ languages with Markdown, HTML or JSON output and layout. Free for research and small teams, not for rivals. Parameters: 5.3B; Context window: 256K; License: OpenRAIL-M (Datalab); Released: 2026-03-18. - [Nemotron 3 Nano 4B](https://ai-opensource.com/models/nemotron-3-nano-4b/): NVIDIA's 4B edge model, a Mamba-Transformer hybrid compressed from Nano 9B v2, with switchable reasoning for RTX PCs and Jetson. Parameters: 3.97B; Context window: 256K; License: NVIDIA Nemotron; Released: 2026-03-16. - [Nemotron 3](https://ai-opensource.com/models/nemotron-3/): NVIDIA's open hybrid Mamba-Transformer MoE pair: Nano 30B-A3B and Super 120B-A12B, 1M context, permissive license, open data. Parameters: 31.6B–120B (12B active); Context window: 1M; License: NVIDIA Nemotron; Released: 2026-03-11. - [S2 Pro](https://ai-opensource.com/models/s2-pro/): Fish Audio's 4.6B text-to-speech model with free-form inline tags and voice cloning from 10 to 30 seconds. Research and non-commercial only. Parameters: 4.56B; License: Fish Audio Research; Released: 2026-03-09. - [Sarvam 30B and 105B](https://ai-opensource.com/models/sarvam-30b-105b/): India's from-scratch open reasoning models: a 105B mixture of experts (10.3B active) and a 30B (2.4B active), 128K context, Apache 2.0. Parameters: 32.1B–106B (10.3B active); Context window: 128K; License: Apache 2.0; Released: 2026-03-06. - [LTX-2.3](https://ai-opensource.com/models/ltx-2-3/): LTX's previous-generation 22B video-and-audio model, still among the most downloaded video models. Free below US$10M revenue. Parameters: 22B; License: LTX-2 Community; Released: 2026-03-05. - [Jina Embeddings v5](https://ai-opensource.com/models/jina-embeddings-v5/): Jina AI's v5 embedding pair: 677M and 239M multilingual models distilled from Qwen3-Embedding-4B, free for non-commercial use only. Parameters: 239M–677M; Context window: 32K; License: CC BY-NC; Released: 2026-02-18. - [Qwen3.5](https://ai-opensource.com/models/qwen-3-5/): Alibaba's Apache 2.0 family from 0.8B to 397B: image and video input, 201 languages, 256K context and a hybrid linear-attention design. Parameters: 873M–403B (17B active); Context window: 256K; License: Apache 2.0; Released: 2026-02-16. - [GLM-5](https://ai-opensource.com/models/glm-5/): Z.ai's February 2026 MIT-licensed 744B-A40B flagship that moved GLM from coding to engineering; it has three newer successors. Parameters: 754B (40B active); Context window: 198K; License: MIT; Released: 2026-02-12. - [MiniMax-M2.5](https://ai-opensource.com/models/minimax-m2-5/): MiniMax's February 2026 230B-A10B coding and agent model under a permissive-style license; M2.7 and M3 have followed. Parameters: 229B (10B active); Context window: 192K; License: MiniMax Model; Released: 2026-02-12. - [MOSS-TTSD](https://ai-opensource.com/models/moss-ttsd/): OpenMOSS's 8B model for spoken dialogue, with up to 5 speakers and 60 minutes in one pass, plus voice cloning. Apache 2.0. Parameters: 8.36B; License: Apache 2.0; Released: 2026-02-10. - [Voxtral Realtime](https://ai-opensource.com/models/voxtral-realtime/): Mistral's Apache 2.0 4B streaming speech-to-text model: 13 languages and a delay you set between 80 ms and 2.4 s. Parameters: 4.43B; Context window: 128K; License: Apache 2.0; Released: 2026-02-04. - [MiniCPM-o 4.5](https://ai-opensource.com/models/minicpm-o-4-5/): OpenBMB's 9B model that sees, listens and speaks at once, with full-duplex live streaming, English and Chinese voices and Apache 2.0. Parameters: 9.37B; Context window: 40K; License: Apache 2.0; Released: 2026-02-03. - [GLM-OCR](https://ai-opensource.com/models/glm-ocr/): Z.ai's 0.9B MIT-licensed document OCR model: 94.62 on OmniDocBench V1.5 as reported, with an SDK, tables, formulas and form extraction. Parameters: 900M; License: MIT; Released: 2026-02-02. - [Step 3.5 Flash](https://ai-opensource.com/models/step-3-5-flash/): StepFun's 196B Apache 2.0 agent model from February 2026: 11B active, 256K context. Legacy, but still widely downloaded. Parameters: 199B (11B active); Context window: 256K; License: Apache 2.0; Released: 2026-02-01. - [Qwen3-ASR](https://ai-opensource.com/models/qwen3-asr/): Alibaba's Apache 2.0 speech-to-text pair (1.7B, 0.6B): 30 languages plus 22 Chinese dialects, singing, streaming and timestamps. Parameters: 918M–2.35B; License: Apache 2.0; Released: 2026-01-29. - [DeepSeek-OCR 2](https://ai-opensource.com/models/deepseek-ocr-2/): DeepSeek's 3.4B Apache 2.0 document OCR model, whose new encoder reorders visual tokens by meaning instead of reading in scan order. Parameters: 3.39B; License: Apache 2.0; Released: 2026-01-27. - [Kimi K2.5](https://ai-opensource.com/models/kimi-k2-5/): Moonshot's January 2026 1T-parameter native multimodal agent model with Agent Swarm: 32B active, 256K context, modified MIT. Parameters: 1.03T (32B active); Context window: 256K; License: Modified MIT; Released: 2026-01-27. - [ACE-Step 1.5](https://ai-opensource.com/models/ace-step-1-5/): An MIT-licensed music foundation model: full songs in seconds, from about 4 GB of VRAM, with covers, repainting and 50+ languages. Parameters: 2B–4B; License: MIT; Released: 2026-01-23. - [Qwen3-TTS](https://ai-opensource.com/models/qwen3-tts/): Alibaba's Apache 2.0 text-to-speech models in 1.7B and 0.6B sizes: 10 languages, 3-second voice cloning, streaming at about 97 ms. Parameters: 910M–1.92B; License: Apache 2.0; Released: 2026-01-22. - [VibeVoice-ASR](https://ai-opensource.com/models/vibevoice-asr/): Microsoft's MIT-licensed 8.7B speech-to-text model: an hour of audio in one pass with speaker labels, timestamps and 50+ languages. Parameters: 2.81B–8.67B; Context window: 64K; License: MIT; Released: 2026-01-21. - [GLM-4.7-Flash](https://ai-opensource.com/models/glm-4-7-flash/): Z.ai's MIT-licensed 30B-A3B model for local coding and agents, from January 2026 and still downloaded about a million times a month. Parameters: 31.2B (3B active); Context window: 198K; License: MIT; Released: 2026-01-19. - [TranslateGemma](https://ai-opensource.com/models/translategemma/): Google's Gemma 3-based translation models in 4B, 12B and 27B sizes for 55 languages, which also translate text inside images. Parameters: 4.3B–27.4B; Context window: 2K; License: Gemma Terms; Released: 2026-01-15. - [voyage-4-nano](https://ai-opensource.com/models/voyage-4-nano/): Voyage AI's first open-weight embedding model: 346M parameters, Apache 2.0, with vectors shared with the hosted Voyage 4 models. Parameters: 346M; Context window: 32K; License: Apache 2.0; Released: 2026-01-15. - [HeartMuLa](https://ai-opensource.com/models/heartmula/): An Apache 2.0 open song model from the HeartMuLa team: lyrics and style tags in, songs of up to four minutes out. Parameters: 3.94B; License: Apache 2.0; Released: 2026-01-14. - [pplx-embed](https://ai-opensource.com/models/pplx-embed/): Perplexity's MIT-licensed embedding family: dense, contextual and late-interaction models, from 0.6B to 9B, with images in v2. Parameters: 590M–8.4B; Context window: 32K; License: MIT; Released: 2026-01-14. - [Pocket TTS](https://ai-opensource.com/models/pocket-tts/): Kyutai's 100M voice-cloning speech model that runs faster than real time on a laptop CPU. Six languages, CC BY 4.0 weights behind a gate. Parameters: 18M–100M; License: CC BY 4.0; Released: 2026-01-13. - [Qwen3-VL-Reranker](https://ai-opensource.com/models/qwen3-vl-reranker/): Alibaba's Apache 2.0 multimodal rerankers (2B and 8B) that score text, images, screenshots and video against a query. Parameters: 2.13B–8.77B; Context window: 32K; License: Apache 2.0; Released: 2026-01-08. - [Qwen3-VL-Embedding](https://ai-opensource.com/models/qwen3-vl-embedding/): Alibaba's Apache 2.0 multimodal embedding models (2B and 8B) that put text, images, screenshots and video in one search space. Parameters: 2.13B–8.14B; Context window: 32K; License: Apache 2.0; Released: 2026-01. - [Fun-CosyVoice 3](https://ai-opensource.com/models/fun-cosyvoice3/): Alibaba Tongyi's Apache 2.0 0.5B zero-shot TTS: 9 languages, 18+ Chinese dialects, voice cloning and streaming from 150 ms. Parameters: 500M; License: Apache 2.0; Released: 2025-12-15. - [Olmo 3.1](https://ai-opensource.com/models/olmo-3-1/): Ai2's fully open 32B and 7B models: weights, training data, code and checkpoints all public, Apache 2.0. Parameters: 7.3B–32.2B; Context window: 64K; License: Apache 2.0; Released: 2025-12-12. - [DeepSeek-V3.2](https://ai-opensource.com/models/deepseek-v3-2/): DeepSeek's MIT-licensed 685B reasoning model with sparse attention for cheaper long context; the V4 family has since replaced it. Parameters: 685B (37B active); Context window: 160K; License: MIT; Released: 2025-12-01. - [Z-Image](https://ai-opensource.com/models/z-image/): Alibaba Tongyi's 6B Apache 2.0 image model; Turbo draws photoreal images with English and Chinese text in 8 steps on a 16 GB GPU. Parameters: 6.15B; License: Apache 2.0; Released: 2025-11-26. - [GigaAM-v3](https://ai-opensource.com/models/gigaam-v3/): Sber's 220M to 240M Russian speech recognizer, with punctuated output in its e2e versions. MIT license. Parameters: 220M–240M; License: MIT; Released: 2025-11-19. - [DeepSeek-OCR](https://ai-opensource.com/models/deepseek-ocr/): DeepSeek's 3B MIT-licensed OCR model that packs page text into few vision tokens: about 97% decoding precision at under 10x compression. Parameters: 3.34B (570M active); License: MIT; Released: 2025-10-20. - [Qwen3-VL](https://ai-opensource.com/models/qwen3-vl/): Alibaba's Apache 2.0 vision-language family from 2B to 235B with 256K context. Superseded by Qwen3.5 but still widely used. Parameters: 2.13B–236B (22B active); Context window: 256K; License: Apache 2.0; Released: 2025-09-23. - [Qwen3-Omni](https://ai-opensource.com/models/qwen3-omni/): Alibaba's 30B-A3B omni model that reads text, images, audio and video and answers in text or speech in 10 languages. Apache 2.0. Parameters: 31.7B–35.3B (3B active); Context window: 64K; License: Apache 2.0; Released: 2025-09-22. - [VibeVoice-TTS](https://ai-opensource.com/models/vibevoice-tts/): Microsoft's MIT-licensed VibeVoice TTS: a 1.5B model for up to 90 minutes and 4 speakers, plus a 0.5B realtime model. Research use. Parameters: 1.02B–2.7B; Context window: 64K; License: MIT; Released: 2025-08-25. - [Parakeet TDT 0.6B v3](https://ai-opensource.com/models/parakeet-tdt-0-6b-v3/): NVIDIA's compact multilingual speech-to-text model: 25 European languages in 0.6B parameters, with timestamps and a CC BY 4.0 license. Parameters: 627M; License: CC BY 4.0; Released: 2025-08-14. - [gpt-oss](https://ai-opensource.com/models/gpt-oss/): OpenAI's Apache 2.0 reasoning models from 2025: 117B and 21B mixtures of experts that fit one 80 GB GPU or a 16 GB machine. Parameters: 21B–117B (5.1B active); Context window: 128K; License: Apache 2.0; Released: 2025-08-05. - [Wan 2.2](https://ai-opensource.com/models/wan-2-2/): Alibaba's Apache 2.0 video models: a two-expert 27B text and image to video family, plus a 5B model for a 24 GB GPU. Parameters: 5B–27B (14B active); License: Apache 2.0; Released: 2025-07-28. - [Qwen3-Embedding](https://ai-opensource.com/models/qwen3-embedding/): Alibaba's Apache 2.0 text embedding models in 0.6B, 4B and 8B sizes: 100+ languages, 32K context, adjustable vector length. Parameters: 600M–7.57B; Context window: 32K; License: Apache 2.0; Released: 2025-06-05. - [Qwen3-Reranker](https://ai-opensource.com/models/qwen3-reranker/): Alibaba's Apache 2.0 instruction-aware rerankers (0.6B, 4B, 8B): 100+ languages, 32K context, built on Qwen3. Parameters: 596M–8.19B; Context window: 32K; License: Apache 2.0; Released: 2025-06-05. - [Qwen3](https://ai-opensource.com/models/qwen3/): Alibaba's 2025 Apache 2.0 family from 0.6B to 235B with a switchable thinking mode. Legacy, yet still hugely downloaded. Parameters: 752M–235B (22B active); Context window: 32K; License: Apache 2.0; Released: 2025-04-29. - [Kokoro](https://ai-opensource.com/models/kokoro/): An 82M-parameter Apache 2.0 speech model with 54 preset voices in 8 languages. Hugely popular and fast, but it cannot clone voices. Parameters: 82M; License: Apache 2.0; Released: 2025-01-27. - [Whisper large-v3-turbo](https://ai-opensource.com/models/whisper-large-v3-turbo/): OpenAI's MIT-licensed speech-to-text model: the turbo version is about 8x faster than large and runs in roughly 6 GB of GPU memory. Parameters: 810M–1.54B; License: MIT; Released: 2024-10-01. - [BGE Reranker v2 m3](https://ai-opensource.com/models/bge-reranker-v2-m3/): BAAI's Apache 2.0 multilingual reranker (568M parameters), by far the most downloaded one we list, built on BGE-M3. Parameters: 568M; Context window: 8K; License: Apache 2.0; Released: 2024-03-18. - [BGE-M3](https://ai-opensource.com/models/bge-m3/): BAAI's MIT-licensed embedding model for 100+ languages and 8,192-token inputs, with dense, sparse and multi-vector search in one model. Parameters: 568M; Context window: 8K; License: MIT; Released: 2024-01-30. ## Best of - [Best open-source LLMs in 2026](https://ai-opensource.com/best/best-open-source-llms/): The leading open-weights language models in November 2026, with size, context window, license and the hardware each one needs. - [Best small open models you can run on a laptop](https://ai-opensource.com/best/small-models-for-laptops/): Open-weights models from 1B to 30B parameters that run on a laptop or one gaming GPU, with their memory needs and licenses. - [Best open-source models for coding and AI agents](https://ai-opensource.com/best/coding-and-agents/): Open-weights models built for coding and agent work, from a 121M tool-calling model that runs on a phone to trillion-parameter systems, with licenses and memory needs. - [Best open-source image and video generation models](https://ai-opensource.com/best/image-and-video/): Open-weights image and video generators you can run on your own GPU, with memory needs and licenses compared, from FLUX to Wan and LTX. - [Best open-source speech models for transcription and text-to-speech](https://ai-opensource.com/best/speech-models/): Open-weights speech-to-text and text-to-speech models you can run yourself, with languages, memory needs and licenses compared. - [Open-source AI models by VRAM and hardware](https://ai-opensource.com/best/models-by-vram/): Find an open AI model that fits your hardware: models grouped by the memory they need, from laptops and gaming GPUs to servers and clusters. - [Open-source AI models you can use commercially](https://ai-opensource.com/best/commercial-use/): Open-weights models grouped by license: permissive, conditional and non-commercial. See which can go into a product, and what each one asks of you. - [Best open-source OCR and document models in 2026](https://ai-opensource.com/best/ocr-and-document-models/): The open OCR and document-parsing models of November 2026, from 0.8B on-device models to 5B multilingual parsers, with their licenses and what each is built for. - [Best open-source text-to-speech models in 2026](https://ai-opensource.com/best/text-to-speech-models/): The open text-to-speech models of November 2026: voice cloning, long-form speech, streaming and tiny on-device models, with a clear note on which licenses allow commercial use. - [Best open-source speech recognition (speech-to-text) models in 2026](https://ai-opensource.com/best/speech-recognition-models/): The open speech-to-text models of November 2026, from Whisper to streaming, long-audio and tiny on-device models, with languages, delay and license for each. - [Best open-source embedding models in 2026](https://ai-opensource.com/best/embedding-models/): The open embedding models of November 2026 for search and RAG: text-only and multimodal, from 47M to 27B parameters, with languages, context length and license. - [Best open-source multimodal models in 2026 (images, video and audio)](https://ai-opensource.com/best/multimodal-models/): Open models that read images, video and audio as well as text, from a 1B on-device model to trillion-parameter systems, with the license and context length of each. - [Open-weight AI models made in Europe](https://ai-opensource.com/best/open-models-made-in-europe/): Open-weight models from developers in France, Germany, Switzerland and the Netherlands, across language, image, speech, OCR and embeddings, with the license of each. - [Open-weight AI models made in South Korea](https://ai-opensource.com/best/open-models-made-in-korea/): Open-weight models from South Korean developers: the four language models built for the government's sovereign AI project, KRAFTON's two Raon models and an on-device speech model. - [Best open-source translation models in 2026](https://ai-opensource.com/best/translation-models/): The open models for translation in November 2026: text translators from 0.8B to 36B, one that also reads text in images, and speech models that translate audio, with their licenses. - [Best open-source reranker models in 2026](https://ai-opensource.com/best/reranker-models/): The open reranker models of November 2026 for search and RAG: text-only and multimodal, from 0.6B to 8B parameters, with languages, context length, license and size. - [Best open-source music and sound-generation models in 2026](https://ai-opensource.com/best/music-and-sound-generation/): The open music and sound models of November 2026: songs with vocals, a live instrument and sound effects, with the hardware and license limits spelled out. ## Compare models - [ACE-Step 1.5 vs MiniMax Music 3: which open song model?](https://ai-opensource.com/compare/ace-step-1-5-vs-minimax-music-3/): ACE-Step 1.5 is an MIT-licensed song model that runs from about 4 GB of VRAM. MiniMax Music 3 is an 11.1B model that writes songs of up to five minutes and needs a 24 GB GPU. - [Apertus 1.5 vs OLMo 3.1: which fully open model should you use?](https://ai-opensource.com/compare/apertus-1-5-vs-olmo-3-1/): Apertus 1.5 reads text, images and audio with a 262K context; OLMo 3.1 is text-only with 64K and has a 32B model. Both are fully open under Apache 2.0. - [BGE-M3 vs Granite Embedding R2: which embedding model should you use?](https://ai-opensource.com/compare/bge-m3-vs-granite-embedding-r2/): BGE-M3 is one 568M multilingual model with dense, sparse and multi-vector search; Granite Embedding R2 is four small models up to 32K context. Compare size and fit. - [Cosmos 3 vs Wan 2.2: which open video model to pick?](https://ai-opensource.com/compare/cosmos-3-vs-wan-2-2/): Cosmos 3 is NVIDIA's world model for robots and driving, with a 4B Edge size; Wan 2.2 is Alibaba's Apache 2.0 text and image to video family with a 5B model for 24 GB GPUs. - [DeepSeek V4 vs Hy3: which large open model should you use?](https://ai-opensource.com/compare/deepseek-v4-vs-hy3/): DeepSeek V4 is MIT-licensed with a 1M-token context in 1.6T and 284B sizes; Hy3 is one 295B Apache 2.0 model with 256K. Compare size, context and license. - [FLUX 3 Klein vs Krea 2: which image model to pick?](https://ai-opensource.com/compare/flux-3-image-klein-vs-krea-2/): FLUX 3 Image Klein edits images and fits in 8 GB with INT8, while Krea 2 draws varied images from text only. Both are gated, with revenue caps of US$5 million and US$1 million. - [FLUX 3 Image Klein vs Qwen-Image-2.1: licenses, memory and features](https://ai-opensource.com/compare/flux-3-image-klein-vs-qwen-image-2-1/): FLUX 3 Image Klein and Qwen-Image-2.1 both generate and edit images from references. Compare their licenses, memory needs and what each does best. - [FLUX.3 Image Klein vs Z-Image: which open image model should you use?](https://ai-opensource.com/compare/flux-3-image-klein-vs-z-image/): FLUX.3 Image Klein edits images from up to 10 references under a revenue-limited license; Z-Image is text-to-image under Apache 2.0. Compare editing, license and memory. - [Gemma 4 vs Qwen3.8-27B: which open model should you run on one GPU?](https://ai-opensource.com/compare/gemma-4-vs-qwen-3-8-27b/): Gemma 4 from Google and Qwen3.8-27B from Alibaba are Apache 2.0 models that fit one GPU. Compare sizes, input types, memory needs and where each fits. - [GLM-5.3 Flash vs MiniMax M3: which multimodal model should you use?](https://ai-opensource.com/compare/glm-5-3-flash-vs-minimax-m3/): GLM-5.3 Flash is 320B, MIT-licensed and built for cheap long context; MiniMax M3 is 428B with reasoning modes and a community license. Compare size and license. - [GLM-5.3 vs DeepSeek V4.1 Flash: open models for coding agents compared](https://ai-opensource.com/compare/glm-5-3-vs-deepseek-v4-1-flash/): GLM-5.3 from Z.ai and DeepSeek V4.1 Flash are open-weights models for coding and agents with a 1M-token context. Compare size, memory needs and licenses. - [GLM-OCR vs PaddleOCR-VL: which small document parser should you use?](https://ai-opensource.com/compare/glm-ocr-vs-paddleocr-vl/): Two 0.9B document parsers: GLM-OCR (MIT, SDK, form extraction) and PaddleOCR-VL (Apache 2.0, charts and seals). Compare tooling, license and scores. - [Index-Translate vs TranslateGemma: which one to pick?](https://ai-opensource.com/compare/index-translate-vs-translategemma/): Index-Translate covers 150 languages under Apache 2.0 and follows glossary and tone instructions. TranslateGemma covers 55 under the Gemma Terms and has far more downloads. We found no independent comparison. - [K-EXAONE 2.0 vs A.X K2: which Korean sovereign AI model should you run?](https://ai-opensource.com/compare/k-exaone-2-0-vs-a-x-k2/): K-EXAONE 2.0 (750B, ten languages, 262K context) or A.X K2 (688B, five languages, strongest at math and Korean)? Both are Apache 2.0. Compare size, hardware and fit. - [Kimi K3.1 Pro vs Qwen3.8-Max: which giant open model fits you?](https://ai-opensource.com/compare/kimi-k3-1-pro-vs-qwen-3-8-max/): Kimi K3.1 Pro and Qwen3.8-Max are the largest open-weights models you can download. Compare size, input types, memory needs and licenses side by side. - [Kokoro vs Pocket TTS: which tiny open TTS model should you use?](https://ai-opensource.com/compare/kokoro-vs-pocket-tts/): Kokoro has 54 fixed voices in 8 languages under Apache 2.0; Pocket TTS clones a voice from 5 seconds of audio in 6 languages under CC BY 4.0. Both run on a CPU. - [LTX-2.5 vs Wan 2.2: open-source video generation compared](https://ai-opensource.com/compare/ltx-2-5-vs-wan-2-2/): LTX-2.5 makes video with synchronized sound, while Wan 2.2 is Apache 2.0 and runs on a 24 GB GPU. Compare memory needs, licenses and features. - [MiniCPM-OCR vs LightOnOCR-3: which Apache 2.0 OCR model should you use?](https://ai-opensource.com/compare/minicpm-ocr-vs-lightonocr-3/): MiniCPM-OCR is one 1.3B model for Chinese and English pages; LightOnOCR-3 offers 4B, 1B and 0.8B sizes with boxes and chart data. Compare tooling and fit. - [Mistral Large 4 vs Muse Spark Open: two giant open-weights models](https://ai-opensource.com/compare/mistral-large-4-vs-muse-spark-open/): Mistral Large 4 and Muse Spark Open were both released in October 2026. Compare size, input types, memory needs and the license conditions of each. - [MOSS-Image vs Z-Image: which Apache 2.0 image model?](https://ai-opensource.com/compare/moss-image-vs-z-image/): MOSS-Image draws long English and Chinese text and follows JSON layouts; Z-Image Turbo makes photoreal images in 8 steps on a 16 GB GPU. Both are Apache 2.0. - [Nemotron 3.5 Lightning vs Qwen3.6: which 3B-active model to use?](https://ai-opensource.com/compare/nemotron-3-5-lightning-vs-qwen-3-6/): Nemotron 3.5 Lightning is a 30B-A3B hybrid for one H100 with 1M context; Qwen3.6 is a 35B-A3B plus 27B dense pair with image input. Compare fit and results. - [Nemotron 3 Nano Omni vs MiniCPM-o 4.5: which open omni model?](https://ai-opensource.com/compare/nemotron-3-nano-omni-vs-minicpm-o-4-5/): NVIDIA's 31B omni model reads video, audio and images and answers in text. OpenBMB's 9B MiniCPM-o 4.5 also speaks. Compare memory, languages and license. - [OmniVoice vs VoxCPM2: which open text-to-speech model should you use?](https://ai-opensource.com/compare/omnivoice-vs-voxcpm2/): OmniVoice covers 600+ languages but its weights are non-commercial; VoxCPM2 covers 30 languages under Apache 2.0. Compare languages, voice design and license. - [Ornith 1.5 vs Qwen3.6: which agentic coding model should you use?](https://ai-opensource.com/compare/ornith-1-5-vs-qwen-3-6/): Ornith 1.5 is an MIT-licensed Qwen3.5 derivative in four sizes; Qwen3.6 is Alibaba's Apache 2.0 pair with image input. Compare sizes, licenses and evidence. - [Parakeet TDT 0.6B v3 vs Cohere Transcribe: which should you use?](https://ai-opensource.com/compare/parakeet-tdt-0-6b-v3-vs-cohere-transcribe/): Parakeet v3 covers 25 European languages with timestamps under CC BY 4.0; Cohere Transcribe covers 14 languages including Chinese, Japanese and Korean under Apache 2.0. - [PLaMo 3 Translate vs Hy-MT2: which open translation model?](https://ai-opensource.com/compare/plamo-3-translate-vs-hy-mt2/): PLaMo 3 Translate covers 53 languages under a revenue-limited license, while Hy-MT2 covers 33 under plain Apache 2.0 and has a far larger community. PFN's own charts favor PLaMo. - [Qwen3-ASR vs Voxtral Realtime: which streaming model should you use?](https://ai-opensource.com/compare/qwen3-asr-vs-voxtral-realtime/): Qwen3-ASR covers 30 languages plus 22 Chinese dialects with timestamps; Voxtral Realtime is one 4B model with an 80 ms to 2.4 s delay. Compare languages and latency. - [Qwen3-Embedding vs EmbeddingGemma 2: which should you use?](https://ai-opensource.com/compare/qwen3-embedding-vs-embeddinggemma-2/): Qwen3-Embedding is text and code in three sizes; EmbeddingGemma 2 puts text, images, audio and video in one space. Compare context, size and licenses. - [Ripple-65B vs Nanbeige4.3-33B: which new agent model?](https://ai-opensource.com/compare/ripple-65b-vs-nanbeige-4-3-33b/): Ripple-65B Preview is a dense 65B model with a 1M context under Apache 2.0; Nanbeige4.3-33B is a 33B model with a 512K context and a 23 GB INT4 build. - [S2 Pro vs Higgs TTS 3: which non-commercial TTS model?](https://ai-opensource.com/compare/s2-pro-vs-higgs-tts-3/): Fish Audio's S2 Pro and Boson's Higgs TTS 3 are both 4.6B text-to-speech models with inline tags and cloning, and both are non-commercial. Compare licenses and hardware. - [Solar Open 2 vs Motif 3: which Korean agent model should you run?](https://ai-opensource.com/compare/solar-open-2-vs-motif-3/): Solar Open 2 has 250B parameters, a 1M context and a naming rule; Motif 3 has 314B, a 256K context and the plain MIT license. Compare hardware, languages and licenses. - [Stable Audio 3 vs MOSS-SoundEffect v2.0: which open sound-effect model?](https://ai-opensource.com/compare/stable-audio-3-vs-moss-soundeffect-v2/): Stable Audio 3 makes music and sound effects of up to 6 min 20 s in eight steps, free below US$1M revenue. MOSS-SoundEffect v2.0 makes 30-second effects under Apache 2.0. - [Whisper large-v3-turbo vs Nemotron 3.5 Transcribe: which to use?](https://ai-opensource.com/compare/whisper-large-v3-turbo-vs-nemotron-3-5-transcribe/): Whisper large-v3-turbo is small and widely supported; Nemotron 3.5 Transcribe adds speaker labels and prompt-controlled style. Compare memory, languages and licenses. - [ZONOS3-Mini vs Drama 3: which new open TTS model?](https://ai-opensource.com/compare/zonos3-mini-vs-drama-3/): ZONOS3-Mini is an 800M Apache 2.0 model for 43 languages that runs in a browser. Drama 3 is a 36B model for 83 languages, free only below US$500K revenue. ## Guides - [What is an open-source AI model? A plain-English guide](https://ai-opensource.com/guides/what-is-an-open-source-ai-model/): An open AI model is one you can download and run yourself. Learn what open weights and open source mean, what you can do with these models and how to start safely. - [Open weights vs open source AI: what is the difference?](https://ai-opensource.com/guides/open-weights-vs-open-source/): Open weights means a model's trained parameters can be downloaded. Open-source AI, as the OSI defines it, also needs code and training-data information. Here is why it matters. - [AI model licenses explained: can you use an open model commercially?](https://ai-opensource.com/guides/ai-licenses-explained/): A plain-English guide to open AI licenses: Apache 2.0, MIT, community licenses with revenue caps, research-only terms and gated downloads, plus a checklist to run before you ship. - [How much memory do I need to run an AI model? A simple guide](https://ai-opensource.com/guides/how-much-memory-do-i-need/): Estimate the VRAM or RAM an open AI model needs: the parameters-times-bytes rule, what quantization saves, why context length and mixture of experts matter, with worked tables. - [How to run an AI model on your own computer](https://ai-opensource.com/guides/how-to-run-ai-models-locally/): A beginner's walkthrough for running open language models locally with Ollama or LM Studio: check your memory, pick a model, run it and connect other apps to it. - [Quantization explained: Q4, FP8, GGUF and what you give up](https://ai-opensource.com/guides/quantization-explained/): Quantization shrinks a model by storing its weights with fewer bits. Learn what FP8, INT4, Q4_K_M, GGUF, MXFP4 and NVFP4 mean, how much they save and how to pick one. - [How to choose an open AI model: a practical decision guide](https://ai-opensource.com/guides/choosing-an-open-model/): Choose an open model by task, hardware, license and capabilities. A step-by-step decision guide with shortlists for chat, coding, documents, images, video, speech, music, translation and search. - [Fine-tuning an open model for beginners: LoRA, QLoRA and when you need it](https://ai-opensource.com/guides/fine-tuning-for-beginners/): Learn when to fine-tune an open AI model instead of prompting or using RAG, how LoRA and QLoRA cut the cost, what data you need and what to check in the license. - [How to evaluate AI models yourself: look past the benchmark charts](https://ai-opensource.com/guides/how-to-evaluate-models-yourself/): Benchmark scores in model announcements are mostly self-reported. Learn how to build a small test set, compare models fairly and judge quality, speed, memory and license for your task. - [Running image and video AI models locally: ComfyUI, memory and licenses](https://ai-opensource.com/guides/running-image-and-video-models-locally/): How to run open image and video generators such as FLUX, Z-Image, Krea 2, MOSS-Image, Anima, Wan, LTX and Cosmos 3 on your own GPU: ComfyUI basics, memory, offloading, LoRAs and license checks. - [How to choose an open-source OCR model for PDFs and scans](https://ai-opensource.com/guides/how-to-choose-an-ocr-model/): Compare open OCR and document-parsing models by size, license, language and tooling, and find out which one fits invoices, long PDFs, Chinese pages or a laptop. - [How to choose an open-source embedding model for search and RAG](https://ai-opensource.com/guides/how-to-choose-an-embedding-model/): Pick an open embedding model by language, input length, modality, size and license, and compare BGE-M3, Qwen3-Embedding, EmbeddingGemma 2, Granite Embedding R2 and more. - [How to choose an open-source text-to-speech model](https://ai-opensource.com/guides/how-to-choose-a-text-to-speech-model/): Compare open text-to-speech models by languages, latency, voice cloning and license: Qwen3-TTS, VoxCPM2, Fun-CosyVoice3, OmniVoice, Voxtral TTS, Chatterbox and tiny CPU models like Kokoro. - [How to choose an open-source speech-to-text model](https://ai-opensource.com/guides/how-to-choose-a-speech-to-text-model/): Compare open speech-to-text models by languages, live streaming, speaker labels, memory and license: Whisper large-v3-turbo, Qwen3-ASR, Voxtral Realtime, VibeVoice-ASR and more. - [How to choose a runner or inference engine for open models](https://ai-opensource.com/guides/choosing-an-inference-engine/): Ollama, LM Studio, llama.cpp, MLX-LM, vLLM, SGLang, TensorRT-LLM and KTransformers compared: which tool fits a laptop, a Mac, a phone, a home server or a production API. - [Self-hosting an open model vs using an API: how to decide](https://ai-opensource.com/guides/self-hosting-vs-api/): Should you run an open model yourself or call a hosted API? Compare privacy, cost structure, hardware, licenses and maintenance, with a checklist and real memory figures. - [How to pick an open model for a coding agent](https://ai-opensource.com/guides/open-models-for-coding-agents/): Choose an open model for agentic coding: what to check (tool calling, context, memory, license), which models fit one GPU or a server, and how to test them on your own repository. - [How to choose an open-source translation model](https://ai-opensource.com/guides/how-to-choose-a-translation-model/): Compare open translation models by language coverage, license, size and instruction support, and find which one fits documents, subtitles, Japanese, images or a laptop. - [How to choose an open-source music generation model](https://ai-opensource.com/guides/how-to-choose-a-music-generation-model/): Compare open music and sound models by what they make, hardware and license: MiniMax Music 3, YuE2, ACE-Step 1.5, HeartMuLa, Stable Audio 3, Magenta RealTime 2 and MOSS-SoundEffect. ## Licenses - [Apache License 2.0](https://ai-opensource.com/licenses/apache-2-0/): One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions. - [bilibili Model Use License Agreement](https://ai-opensource.com/licenses/bilibili-model-use/): Bilibili's license for IndexTTS, which the IndexTTS-2.5 card and the GitHub repository both point to. It lets you use, modify, train on, distribute and build derivatives of the model, worldwide and royalty free. The text has no general ban on commercial use, but a user or affiliate above 100 million monthly active users or RMB 1 billion in annual revenue needs a separate written license. - [Boson Higgs TTS 3 Research and Non-Commercial License](https://ai-opensource.com/licenses/boson-higgs-non-commercial/): Boson AI's license for the Higgs TTS 3 weights, last updated on 8 July 2026. It calls itself a source-available license, not an open-source one. Research and non-commercial use is free, and any commercial purpose needs a separate written license from Boson. The one exception is a Creator Use Grant, which lets digital creators make and monetize podcasts, videos, audiobooks and social posts with the model's audio if they credit Boson AI's Higgs Audio. - [BreezeBlue Research and Non-Commercial License Agreement v1.1](https://ai-opensource.com/licenses/breezeblue-non-commercial/): The license for the Breeze TTS 2 weights, version 1.1 of 1 September 2026, issued by Resonia, Inc., which does business as BreezeBlue. It is free for research and non-commercial use and grants no commercial rights at all, with no revenue-threshold or creator exception. Audio you make with the open weights carries the same limits. - [Creative Commons Attribution 4.0 (CC BY 4.0)](https://ai-opensource.com/licenses/cc-by-4-0/): A Creative Commons license that lets anyone copy, share and adapt the work for any purpose, including commercial use, as long as they give credit. It was written for creative and data works rather than software, but some speech and data releases use it for model weights. - [Creative Commons Attribution-NonCommercial (CC BY-NC)](https://ai-opensource.com/licenses/cc-by-nc/): A Creative Commons license that lets anyone copy, share and adapt the work with credit, but only for non-commercial purposes. Some model cards name it without a version number, and the current version is 4.0. - [CircleStone Labs Non-Commercial License v1.2](https://ai-opensource.com/licenses/circlestone-non-commercial/): The license for CircleStone Labs' Anima image models. The weights are free only for non-commercial use, but the images you generate can be used commercially. Hosting the model for payment, or building it into a monetized product, needs a separate commercial license from CircleStone Labs. - [AI Pubs Open RAIL-M License (Modified, Datalab)](https://ai-opensource.com/licenses/datalab-openrail-m/): A modified Open RAIL-M license. Free for research and personal use, and for commercial use only by organizations with under US$2 million in prior-year revenue and under US$2 million raised in total. It also bans using the model to compete with Datalab and requires derivatives and outputs to carry the same license. - [AI Pubs Open RAIL-M License (Modified, Datalab Surya)](https://ai-opensource.com/licenses/datalab-surya-openrail-m/): The same modified Open RAIL-M license as Datalab's Chandra models, with higher limits. Free for research and personal use, and for commercial use only by organizations with under US$5 million in prior-year revenue and under US$5 million raised in total. It also bans using the model to compete with Datalab and requires derivatives and outputs to carry the same license. - [ERNIE 5.5 Flash Base Model License Agreement v1.0](https://ai-opensource.com/licenses/ernie-5-5-flash-base/): Free for non-commercial use, and free for commercial use up to US$50 million of related annual revenue if you show 'Powered by ERNIE' and send Baidu a one-time notice, but only outside the EU, the UK and South Korea. Above US$50 million you need Baidu's written authorization. - [Fish Audio Community License Agreement 1.0](https://ai-opensource.com/licenses/fish-audio-community/): Free for research and non-commercial use, and free for commercial use while you and your affiliates earn less than US$500,000 a year. Above that, or to offer the model to others as a text-to-speech, voice-cloning or inference service at any size, you need a separate written license from Fish Audio. You must show "Built with Fish Audio", you may not clone a real person's voice without consent, and you may not use the model or its audio to train other speech or foundation models. - [Fish Audio Research License Agreement](https://ai-opensource.com/licenses/fish-audio-research/): Fish Audio's license for the S2 Pro weights, last updated on 7 March 2026. Research and non-commercial use is free, including hobby projects, evaluation and testing. Any commercial purpose, including your own organization's internal operations, needs a separate written license from Fish Audio. If you share the weights or a derivative, you pass on the agreement, keep a notice crediting 39 AI, Inc. and show "Built with Fish Audio". - [FLUX Kommunity License v1.0](https://ai-opensource.com/licenses/flux-kommunity/): Black Forest Labs' license for its open FLUX 3 weights. It is free for personal and non-commercial use, and free for commercial use only by "Qualifying Users", meaning you and your affiliates earn under US$5 million a year. Larger companies need a commercial license or the paid API. - [Gemma Terms of Use](https://ai-opensource.com/licenses/gemma-terms/): Google's custom license for the Gemma models released before Gemma 4, including TranslateGemma. It allows commercial use, changes and redistribution with no revenue cap, but every copy and every derivative must carry the same use restrictions, and Google may restrict use it believes breaks the terms. - [GLM-5.3 License](https://ai-opensource.com/licenses/glm-5-3/): Close to MIT. The only extra condition applies to giant hosting businesses: a model-as-a-service operator with more than US$10 billion of revenue must pass a Z.AI security review. Everyone else can treat it like MIT. - [Ideogram Non-Commercial Model Agreement](https://ai-opensource.com/licenses/ideogram-non-commercial/): Ideogram's license for the Ideogram 4 weights, last updated on 3 June 2026. The weights are free only for non-commercial purposes, which the text defines narrowly. A company may test and prototype but not deploy, and generating images for products that earn revenue is excluded. Commercial use needs a separate agreement that Ideogram may grant or refuse as it likes. - [Kimi K3 License](https://ai-opensource.com/licenses/kimi-k3/): An MIT-style license with two scale conditions. Most companies can use it commercially without asking, but large hosting businesses need a deal and very large products must credit the model. Kimi K3.1 Pro uses the same terms under its own name. - [Krea 2 Community License Agreement v1](https://ai-opensource.com/licenses/krea-2-community/): Krea's license for the Krea 2 Raw and Turbo weights, dated 22 June 2026. You may use, modify and share the models and create derivatives, and commercial use is allowed only while you and your affiliates earn under US$1 million a year. Larger organizations need an enterprise license from Krea. - [LFM Open License v1.0](https://ai-opensource.com/licenses/lfm-open-1-0/): An Apache-2.0-style license from Liquid AI with one big catch. Commercial use is licensed only while your company, together with its affiliates, earns under US$10 million a year. Above that line you are not licensed and need your own agreement. - [LTX-2 Community License Agreement](https://ai-opensource.com/licenses/ltx-2-community/): The custom license for LTX-2 and LTX-2.3, dated 5 January 2026. Organizations with annual revenue under US$10 million can use, modify and host the models for any purpose, subject to a list of banned uses. Organizations at or above that figure need a paid commercial agreement. LTX-2.5 uses the newer LTX-2.x license instead. - [LTX-2.x Community License Agreement](https://ai-opensource.com/licenses/ltx-2x-community/): A custom license for LTX video models. Organizations with annual revenue under US$10 million can use, modify and host the model for any purpose; organizations at or above that figure need a paid commercial license, though they may still evaluate it in non-production settings. - [MiniMax Community License](https://ai-opensource.com/licenses/minimax-community/): A license that grants rights for non-commercial purposes and then lays out the steps for commercial use: show a credit line, send a notice, and get written authorization if your product earns over US$20 million a year. Military use is banned. Get a legal read before shipping a product on it. - [MiniMax H3 Community License Agreement](https://ai-opensource.com/licenses/minimax-h3-community/): A community license that lets you use, modify and distribute MiniMax H3, but only in the "Applicable Territory": the whole world except the European Union, the United Kingdom, South Korea and the United States. Commercial products earning over US$20 million a year need MiniMax's written authorization, every commercial product must show "MiniMax H3" in its interface, and you may not use the model or its outputs to train other AI models. Get legal advice, especially if you or your users are in an excluded region. - [MiniMax-H3-Image Community License Agreement](https://ai-opensource.com/licenses/minimax-h3-image-community/): Built on the MiniMax H3 license, with a much lower commercial threshold. You may use, modify and distribute MiniMax-H3-Image only in the "Applicable Territory" (the world except the European Union, the United Kingdom, South Korea and the United States). You need MiniMax's written authorization if you and your affiliates earn more than US$1 million a year, or if you offer the model as a hosted service to anyone else. Get legal advice before building a business on it. - [MiniMax M2.7 Non-Commercial License](https://ai-opensource.com/licenses/minimax-m2-7-non-commercial/): An MIT-style license that covers non-commercial use only. Personal use, including self-hosting for coding and building your own tools, and non-profit, academic and research use are free. Any commercial use needs separate prior written authorization from MiniMax, and where it is authorized you must show "Built with MiniMax M2.7". - [MiniMax Model License (M2.5)](https://ai-opensource.com/licenses/minimax-model-license/): A permissive-style limited license from MiniMax, with a version date of 13 February 2026. It grants a worldwide, royalty-free right to use, copy, distribute, modify and make derivative works of the model, with no revenue cap in the text. You must pass the agreement on, mark changed files, keep MiniMax's attribution notice and follow a Prohibited Uses Policy. The model card and Hugging Face metadata call it Modified-MIT, but the license file in the repository is titled MiniMax Model License. - [MiniMax-Music3 Community License](https://ai-opensource.com/licenses/minimax-music3-community/): An MIT-style license with commercial conditions. You may use, modify and distribute MiniMax Music 3, but a commercial product must show "MiniMax-Music3" in its interface, and a business earning over US$20 million a year from products that use it needs written authorization. Anyone who offers the model to others must also keep safeguards against infringing outputs. A long acceptable-use policy applies. - [Mistral Ecosystem Open-Weight License](https://ai-opensource.com/licenses/mistral-open-weight/): A modified MIT license. You can use and build on the model freely, but any business that sells hosted access to it needs a separate commercial license from Mistral, with no revenue threshold. - [MIT License](https://ai-opensource.com/licenses/mit/): The shortest and simplest permissive license. You can do almost anything with the model as long as you keep the copyright and license notice. - [Modified MIT License (Moonshot AI)](https://ai-opensource.com/licenses/moonshot-modified-mit/): The standard MIT license with one added sentence. Anyone can use, modify and sell the model, and only very large products (over 100 million monthly active users or over US$20 million of monthly revenue) must show the model's name in their interface. - [Muse Model License](https://ai-opensource.com/licenses/muse-model-license/): Meta's license for Muse Spark Open. Most companies can use it commercially, but a business selling hosted access above US$20 million of revenue needs a separate agreement, and everyone must display "Built with Muse Spark Open". The download is gated. - [NetEase Youdao Model Use License Agreement](https://ai-opensource.com/licenses/netease-youdao-model-use/): The license for NetEase Youdao's Confucius4-R2T2 speech recognizer. Commercial use is licensed unless you or an affiliate had more than 100 million monthly active users or more than RMB 1 billion of annual revenue, and then you need a separate license. It also bars using the model to improve other AI models, apart from non-commercial ones, and prohibits high-risk uses. - [NVIDIA License (non-commercial)](https://ai-opensource.com/licenses/nvidia-license-non-commercial/): NVIDIA's short research-only license, as shipped with LocateAnything-3B. You may use, copy, modify and share the work, but only non-commercially, which the text defines as research or evaluation purposes only. NVIDIA and its affiliates keep the right to use it commercially. - [NVIDIA Nemotron Open Model License](https://ai-opensource.com/licenses/nvidia-nemotron-open-model/): A permissive, Apache-style license from NVIDIA. Commercial use, derivative works and redistribution are allowed, NVIDIA claims no ownership of outputs, and the main duties are to pass the license on and keep notices. - [NVIDIA Open Model Agreement](https://ai-opensource.com/licenses/nvidia-open-model/): A permissive, Apache-style agreement from NVIDIA, released on 2 April 2026. NVIDIA confirms that its works are commercially usable, that you may create and distribute derivative works, and that it claims no ownership of outputs. The main duties are to pass the agreement on and keep notices. - [OpenMDW License Agreement 1.1](https://ai-opensource.com/licenses/openmdw-1-1/): A short permissive license written for machine-learning models by the Linux Foundation and PyTorch Foundation. It covers the model, its data, documentation and software together, and puts no restrictions on what you do with the model's outputs. - [BigScience Open RAIL-M License](https://ai-opensource.com/licenses/openrail-m/): The standard Open RAIL-M license written by BigScience in August 2022. It grants free, worldwide copyright and patent rights, including hosting the model as a service, but ties every copy and every derivative to a list of banned uses such as breaking the law, impersonating people without consent, giving medical advice and fully automated decisions about people's legal rights. - [PLaMo Community License Agreement](https://ai-opensource.com/licenses/plamo-community/): A Japanese community license for the PLaMo models. Personal and academic use is free. Commercial use is free only after you register with Preferred Networks and only while you and your affiliates have at most 1 billion yen of annual revenue or sales; above that you need a separate agreement. If you hand the model or a derived model to others, you must show "Built with PLaMo" and keep "PLaMo" in the names of the models you create. - [Qwen3.8-Max License](https://ai-opensource.com/licenses/qwen-3-8-max/): An MIT-style license from Alibaba's Qwen team for its flagship model. Free for almost everyone, with a name-credit rule for huge products and a separate deal for large hosting and AI work-assistant businesses. - [Qwen Community License 1.0](https://ai-opensource.com/licenses/qwen-community-1-0/): An MIT-style license with a stricter hosting rule than the Qwen3.8-Max License. You can build products with the model freely, but anyone running a model-as-a-service or AI work-assistant business needs a separate license, whatever their revenue. - [Qwen Research License Agreement](https://ai-opensource.com/licenses/qwen-research/): Alibaba's research-only license. You may use, modify and share the model for research or evaluation, but any commercial use needs a separate license from Qwen. Alibaba's hosted API is the usual commercial route. - [Stability AI Community License](https://ai-opensource.com/licenses/stability-ai-community/): Stability AI's license for its open models. Research and non-commercial use is free, and so is commercial use while you and your affiliates earn under US$1 million a year, but commercial users must register with Stability AI and show "Powered by Stability AI". Above US$1 million the license ends and you must ask for an enterprise license. You own the outputs, to the extent the law allows. - [Stellon Labs Community License Agreement](https://ai-opensource.com/licenses/stellon-labs-community/): The license for the KittenTTS 2 weights from KittenML, whose text is dated 5 July 2024. Research, non-commercial and limited commercial use are free. Commercial users must register with Stellon Labs, and the commercial grant ends once you or your affiliates pass US$1 million in annual revenue or in total funding. - [StepFun Model License 1.0](https://ai-opensource.com/licenses/stepfun-model-1-0/): A custom license that lets you use, modify, distribute and sublicense Step 5 Preview free of charge, including commercially. Hardware makers that build the model into phones, vehicles, glasses, wearables, robots or smart-home devices need a separate commercial license, and so does any business that runs a model-as-a-service offering with more than US$20 million of yearly revenue. A product with over 100 million active users, or over US$20 million of revenue in a month, must show "Powered by StepFun AI". Models trained to a material extent on the model's outputs count as derivative works. - [Tencent Hunyuan Community License](https://ai-opensource.com/licenses/tencent-hunyuan-community/): A royalty-free license with real limits. It is valid only outside the European Union, the United Kingdom and South Korea, large services with over 100 million monthly users must ask Tencent first, and you may not use the model or its outputs to improve other AI models. - [Upstage Solar License](https://ai-opensource.com/licenses/upstage-solar/): Apache 2.0 with one added rule. You can use the model commercially and build on it freely, but an AI model you create from it and distribute must have a name that starts with 'Solar' and must show the words 'Built with Solar'. - [webAI Non-Commercial License ver. 1.0](https://ai-opensource.com/licenses/webai-non-commercial/): A short license that webAI, Inc. puts on its TwIL models. It grants a perpetual, worldwide, royalty-free copyright license to use, modify and share the weights, but only for non-commercial research and educational purposes. The same limit follows every derivative work. ## Glossary - [Agent memory](https://ai-opensource.com/glossary/agent-memory/): Agent memory lets an AI agent keep facts, preferences and past experience between sessions instead of starting from a blank context. - [Agentic coding](https://ai-opensource.com/glossary/agentic-coding/): Agentic coding means a model works on software like a developer: it reads the codebase, edits files, runs commands and tests, and keeps going until the task is done. - [AI agent](https://ai-opensource.com/glossary/ai-agent/): An AI agent is a system in which a model plans and carries out multi-step tasks by calling tools, such as running code, searching the web or editing files. - [Base model](https://ai-opensource.com/glossary/base-model/): A base model is the raw result of pretraining, before it has been taught to follow instructions or chat; it is the usual starting point for fine-tuning. - [Benchmark](https://ai-opensource.com/glossary/benchmark/): A benchmark is a standardized test used to compare models, such as a set of maths, coding or knowledge questions; scores are useful but easy to over-read. - [Classifier-free guidance (CFG)](https://ai-opensource.com/glossary/classifier-free-guidance/): Classifier-free guidance, usually called CFG, is a setting in diffusion models that controls how strictly the result follows your prompt, at the price of extra computation. - [Computer-use agent](https://ai-opensource.com/glossary/computer-use-agent/): A computer-use agent looks at screenshots and operates a computer like a person, by clicking, typing and scrolling, and sometimes by running shell commands or code. - [Context window](https://ai-opensource.com/glossary/context-window/): The context window is the maximum amount of text, measured in tokens, that a model can take into account at one time, including your prompt and its own answer. - [Decision model](https://ai-opensource.com/glossary/decision-model/): A decision model reads a situation and a list of questions, then returns a probability for each allowed answer instead of writing text. - [Dense model](https://ai-opensource.com/glossary/dense-model/): A dense model uses all of its parameters to process every token, unlike a mixture-of-experts model, which activates only a part of them. - [Diffusion model](https://ai-opensource.com/glossary/diffusion-model/): A diffusion model generates images or video by starting from random noise and removing the noise step by step until a picture emerges. - [Distillation](https://ai-opensource.com/glossary/distillation/): Distillation trains a smaller student model to imitate a larger teacher model, keeping much of its ability at a fraction of the size. - [Efficient attention (sparse, linear and hybrid)](https://ai-opensource.com/glossary/efficient-attention/): Efficient attention designs let a model read very long inputs with far less memory and compute than standard attention, by looking at fewer tokens or compressing the past. - [Embedding model](https://ai-opensource.com/glossary/embedding-model/): An embedding model turns text, images or other data into a list of numbers that captures its meaning, so that similar items end up close together. - [Fine-tuning](https://ai-opensource.com/glossary/fine-tuning/): Fine-tuning continues a pretrained model's training on your own examples so that it behaves better for a specific task, style or domain. - [Gated model](https://ai-opensource.com/glossary/gated-model/): A gated model is an open model whose download on Hugging Face requires you to log in and accept its terms, or request access, first. - [GGUF](https://ai-opensource.com/glossary/gguf/): GGUF is a single-file model format used by llama.cpp and the tools built on it, which bundles quantized weights with the information needed to run them. - [Hallucination](https://ai-opensource.com/glossary/hallucination/): A hallucination is a confident but false or invented statement produced by an AI model. - [Hugging Face](https://ai-opensource.com/glossary/hugging-face/): Hugging Face is the website where most open models are published: each model has a page with its files, license, model card and download count. - [Inference](https://ai-opensource.com/glossary/inference/): Inference is the act of running a trained model to get an output, such as answering a prompt, as opposed to training it. - [KV cache](https://ai-opensource.com/glossary/kv-cache/): The KV cache is the memory a model uses to remember the earlier tokens of a conversation, and it grows with the length of the context. - [Looped transformer](https://ai-opensource.com/glossary/looped-transformer/): A looped transformer reruns some of its own layers several times for each token, so it gets more computation without storing more parameters. - [LoRA](https://ai-opensource.com/glossary/lora/): LoRA, short for low-rank adaptation, fine-tunes a model by training a small set of added weights instead of changing all of its parameters. - [Machine translation](https://ai-opensource.com/glossary/machine-translation/): Machine translation is software that translates text, speech or text inside images from one language to another. The open models of today usually start from a general model and are then trained only to translate. - [Matryoshka embeddings](https://ai-opensource.com/glossary/matryoshka-embeddings/): Matryoshka embeddings are trained so that the first part of the vector still works on its own, which lets you cut the vector shorter to save storage and search time. - [Mixture of experts (MoE)](https://ai-opensource.com/glossary/mixture-of-experts/): A mixture-of-experts model contains many specialist sub-networks and uses only a few of them for each token, so it can be very large yet fast. - [MTEB (embedding benchmark)](https://ai-opensource.com/glossary/mteb/): MTEB, the Massive Text Embedding Benchmark, scores embedding models on many tasks such as search, classification and clustering, and its public leaderboard is a common way to compare them. - [Multimodal model](https://ai-opensource.com/glossary/multimodal-model/): A multimodal model can handle more than one type of data, such as text, images, audio or video, as input, as output, or both. - [Music generation](https://ai-opensource.com/glossary/music-generation/): Music generation models turn a text description, and often lyrics, into audio: full songs with vocals, instrumental tracks, live streams of music or sound effects. - [Non-commercial license](https://ai-opensource.com/glossary/non-commercial-license/): A non-commercial license lets you use a model for research, study and personal projects, but not to make money from it. - [OCR and document parsing](https://ai-opensource.com/glossary/ocr/): OCR turns an image of text into text; document parsing goes further and rebuilds tables, formulas, charts and reading order, usually as Markdown. - [OmniDocBench](https://ai-opensource.com/glossary/omnidocbench/): OmniDocBench is a benchmark for document parsing: it scores how well a model turns PDF pages of many types into text, tables and formulas. - [On-device AI](https://ai-opensource.com/glossary/on-device-ai/): On-device AI means the model runs on your own phone, laptop or small board instead of in the cloud, so data stays local and it works offline. - [Open-source AI](https://ai-opensource.com/glossary/open-source-ai/): Open-source AI, as defined by the Open Source Initiative, is an AI system you may use, study, modify and share freely, with the weights, the code and detailed information about the training data all available. - [Open weights](https://ai-opensource.com/glossary/open-weights/): An open-weights model is an AI model whose trained parameters, called weights, are published so that anyone can download and run it, usually under a license that sets the terms of use. - [OpenAI-compatible API](https://ai-opensource.com/glossary/openai-compatible-api/): An OpenAI-compatible API is a server that accepts the same requests as the OpenAI API, so apps written for it can use an open model instead. - [OpenRAIL (Responsible AI License)](https://ai-opensource.com/glossary/openrail/): OpenRAIL is a family of licenses that let you use, modify and share a model, including commercially, but require every copy and derivative to carry a list of banned uses. - [Parameters](https://ai-opensource.com/glossary/parameters/): Parameters are the numbers inside a model that are adjusted during training; their count, such as 7B or 70B, is the usual measure of a model's size. - [Permissive license](https://ai-opensource.com/glossary/permissive-license/): A permissive license, such as Apache 2.0 or MIT, lets you use, modify and sell a model with very few conditions, usually just keeping a notice. - [Precision (BF16, FP8, INT4)](https://ai-opensource.com/glossary/precision/): Precision is how many bits are used to store each number in a model, such as 16-bit BF16, 8-bit FP8 or 4-bit INT4; fewer bits mean a smaller, faster model. - [Quantization](https://ai-opensource.com/glossary/quantization/): Quantization stores a model's weights with fewer bits, for example 4 instead of 16, to cut memory use and speed up inference at a small cost in quality. - [Quantization-aware training (QAT)](https://ai-opensource.com/glossary/quantization-aware-training/): Quantization-aware training teaches a model to work with low-precision weights while it is still being trained, so it loses less quality at 4 bits than a model squeezed afterwards. - [Retrieval-augmented generation (RAG)](https://ai-opensource.com/glossary/rag/): Retrieval-augmented generation lets a model answer from your own documents by retrieving the relevant passages and adding them to the prompt. - [Reasoning model](https://ai-opensource.com/glossary/reasoning-model/): A reasoning model works through a problem step by step in a 'thinking' phase, visible or hidden, before it gives its final answer. - [Reinforcement learning (RL)](https://ai-opensource.com/glossary/reinforcement-learning/): Reinforcement learning trains a model by rewarding good outcomes, such as a correct answer or passing tests, instead of copying example text. - [Reranker](https://ai-opensource.com/glossary/reranker/): A reranker re-scores a short list of search results by reading the question and each result together, which is slower but more accurate than embeddings alone. - [Revenue cap (revenue threshold)](https://ai-opensource.com/glossary/revenue-cap/): A revenue cap is a license condition that makes a model free below a certain income and requires a separate agreement above it. - [Sampling steps](https://ai-opensource.com/glossary/sampling-steps/): Sampling steps are the number of passes a diffusion model makes to turn noise into an image or video. Fewer steps mean faster results, usually with some loss of detail or variety. - [Self-hosting](https://ai-opensource.com/glossary/self-hosting/): Self-hosting means running a model on hardware you control, such as your own PC or a rented GPU server, instead of calling a vendor's API. - [Sovereign AI](https://ai-opensource.com/glossary/sovereign-ai/): Sovereign AI means AI that a country or organization can build, host and govern under its own control, from the computing power and data to the model itself. - [Speaker diarization (speaker labels)](https://ai-opensource.com/glossary/speaker-diarization/): Diarization works out who spoke when in a recording and labels each stretch of speech with a speaker, such as Speaker 1 and Speaker 2. - [Speculative decoding and multi-token prediction](https://ai-opensource.com/glossary/speculative-decoding/): Speculative decoding speeds up text generation by letting a cheap draft guess several tokens ahead and having the full model check them all at once. - [Speech recognition (ASR, speech-to-text)](https://ai-opensource.com/glossary/speech-recognition/): Speech recognition, also called ASR or speech-to-text, turns spoken audio into written text. - [Ternary weights (1.58-bit)](https://ai-opensource.com/glossary/ternary-weights/): Ternary weights limit every model weight to one of three values, -1, 0 or +1, so a model needs under 2 bits per weight instead of 16. - [Territory restriction](https://ai-opensource.com/glossary/territory-restriction/): A territory restriction limits where a license is valid: using the model in the excluded countries or regions is not licensed. - [Text encoder](https://ai-opensource.com/glossary/text-encoder/): In an image or video generator, the text encoder is a separate language model that turns your prompt into numbers the generator can use. It adds to the download and to memory use. - [Text-to-speech (TTS)](https://ai-opensource.com/glossary/text-to-speech/): Text-to-speech models turn written text into spoken audio, and many can copy a voice from a short sample. - [Token](https://ai-opensource.com/glossary/token/): A token is a small chunk of text, often a word or part of a word, that a language model reads and writes; models measure length and speed in tokens. - [Tool calling (function calling)](https://ai-opensource.com/glossary/tool-calling/): Tool calling lets a model ask your program to run a function, such as a search or a calculator, and then use the result in its answer. - [Translation metrics](https://ai-opensource.com/glossary/translation-metrics/): Translation metrics are the scores that model cards use to rate a translation. Some count overlap with a human reference, others use a trained model or human reviewers, and their scales cannot be compared. - [Unified memory](https://ai-opensource.com/glossary/unified-memory/): Unified memory is a design, used by Apple silicon and some other chips, in which the CPU and GPU share one pool of memory, so a large model can use most of it. - [Vision-language model (VLM)](https://ai-opensource.com/glossary/vision-language-model/): A vision-language model reads images, and often video, together with text, so you can ask questions about a picture, a screenshot or a document page. - [Voice cloning](https://ai-opensource.com/glossary/voice-cloning/): Voice cloning makes a text-to-speech model speak in a particular person's voice, often from only a few seconds of recording. - [VRAM](https://ai-opensource.com/glossary/vram/): VRAM is the memory on a graphics card; it decides which models a GPU can hold, because the weights must fit in memory for the model to run quickly. - [Word error rate (WER)](https://ai-opensource.com/glossary/word-error-rate/): Word error rate is the standard score for speech recognition: the share of words in a transcript that are wrong, missing or extra compared with a correct reference. ## Tools - [AI Toolkit](https://ai-opensource.com/tools/ai-toolkit/): Ostris's all-in-one training suite for image, video and audio diffusion models, with a web interface and a command line. - [AnythingLLM](https://ai-opensource.com/tools/anythingllm/): An all-in-one app for chatting with your documents, building agents and running local models, as a desktop app or a multi-user Docker server. - [Axolotl](https://ai-opensource.com/tools/axolotl/): A free fine-tuning framework driven by one YAML file, with LoRA, QLoRA, preference tuning and NVFP4 and expert-parallel training for MoE models. - [ComfyUI](https://ai-opensource.com/tools/comfyui/): The node-based studio for open image, video, audio and 3D models: build visual workflows and run them on your own GPU. - [DiffSynth-Studio](https://ai-opensource.com/tools/diffsynth-studio/): The ModelScope community's open-source diffusion engine for running and training image, video and audio models on modest GPUs. - [Diffusers](https://ai-opensource.com/tools/diffusers/): Hugging Face's Python library for running and training diffusion models for images, video and audio. - [ExLlamaV3](https://ai-opensource.com/tools/exllamav3/): A CUDA inference library for consumer GPUs with its own EXL3 quantization, tensor parallelism and CPU offloading for MoE models. - [InvokeAI](https://ai-opensource.com/tools/invokeai/): A self-hosted creative studio for AI image generation, with a layered canvas, node-based workflows and a model manager. - [Jan](https://ai-opensource.com/tools/jan/): A free, open-source desktop alternative to ChatGPT that downloads and runs open models offline, and can also connect to cloud APIs. - [KoboldCpp](https://ai-opensource.com/tools/koboldcpp/): A single-file app that runs GGUF models with a built-in web UI, and also generates images, video, speech and music. - [KTransformers](https://ai-opensource.com/tools/ktransformers/): A research framework that runs very large mixture-of-experts models by sharing the work between a GPU and CPU memory. - [LightX2V](https://ai-opensource.com/tools/lightx2v/): A lightweight inference framework for image and video generation, with distilled models, quantization and many hardware back ends. - [llama.cpp](https://ai-opensource.com/tools/llama-cpp/): The C/C++ engine behind much of local AI: runs quantized models on almost any hardware, with a command-line tool and an OpenAI-compatible server. - [LLaMA-Factory](https://ai-opensource.com/tools/llama-factory/): Fine-tune 100+ models with a zero-code command line or web UI, covering LoRA, QLoRA, DPO and multimodal training. - [LM Studio](https://ai-opensource.com/tools/lm-studio/): A polished desktop app for downloading, chatting with and serving local models, with an agent and offline voice transcription built in. - [LocalAI](https://ai-opensource.com/tools/localai/): A self-hosted server that puts language, vision, voice, image and video models behind OpenAI-, Anthropic- and ElevenLabs-compatible APIs. - [MLC LLM](https://ai-opensource.com/tools/mlc-llm/): A compiler-based engine that deploys the same model on phones, browsers, laptops and GPUs behind an OpenAI-compatible API. - [MLX LM](https://ai-opensource.com/tools/mlx-lm/): Apple's Python package for generating text and fine-tuning language models on Apple silicon Macs with the MLX framework. - [Ollama](https://ai-opensource.com/tools/ollama/): The simplest way to download and run open models on your own computer: one command to chat, plus a local API other apps can use. - [Open WebUI](https://ai-opensource.com/tools/open-webui/): A self-hosted chat interface that works with Ollama and any OpenAI-compatible API, built to run entirely offline. - [SGLang](https://ai-opensource.com/tools/sglang/): A fast open-source serving framework for language, vision and diffusion models, tuned for agents and large-scale deployment. - [stable-diffusion.cpp](https://ai-opensource.com/tools/stable-diffusion-cpp/): A C/C++ engine, in the style of llama.cpp, that runs image and video diffusion models such as FLUX, Qwen-Image, Z-Image and Wan. - [TensorRT LLM](https://ai-opensource.com/tools/tensorrt-llm/): NVIDIA's inference engine for LLMs and visual generation models on NVIDIA GPUs, with optimized kernels and a Python API. - [Text Embeddings Inference](https://ai-opensource.com/tools/text-embeddings-inference/): Hugging Face's fast server for open text-embedding, re-ranker and classification models, shipped as small Docker images. - [Hugging Face Transformers](https://ai-opensource.com/tools/transformers/): The Python library that defines most open models: the reference way to load, run and fine-tune a model straight from its release. - [Unsloth](https://ai-opensource.com/tools/unsloth/): Run and fine-tune open models from a desktop app, a web UI or code, with GGUF export and day-one support for many new releases. - [vLLM](https://ai-opensource.com/tools/vllm/): The most widely used open-source engine for serving language models at scale, with an OpenAI-compatible API. - [vLLM-Omni](https://ai-opensource.com/tools/vllm-omni/): The vLLM project's serving framework for models that take or produce more than text: speech, image, video and action models. - [whisper.cpp](https://ai-opensource.com/tools/whisper-cpp/): A dependency-free C/C++ port of OpenAI's Whisper that transcribes speech on CPUs, Apple Silicon, GPUs and phones. ## Developers - [Abacus.AI](https://ai-opensource.com/developers/abacusai/): A company that sells an AI assistant platform for professionals and enterprises and publishes open Smaug models on Hugging Face, from Smaug-72B in 2024 to the Smaug-Decisions audio, image and text decision models in October 2026, which use Apache 2.0. - [ACE-Step](https://ai-opensource.com/developers/ace-step/): ACE-Step is an open music-generation project co-led by ACE Studio and StepFun. Its ACE-Step 1.5 models, released in January 2026 with a 4B XL version in April, use the MIT license and are trained on licensed, royalty-free and synthetic data. - [Adventists AI](https://ai-opensource.com/developers/adventists-ai/): A voice-agent company, known in Chinese as AI降临派, that releases the Sophon-Realtime full-duplex voice models under Apache 2.0 and the DuplexJev models that decide turn-taking and tone straight from audio under CC BY-NC. - [Allen Institute for AI (Ai2)](https://ai-opensource.com/developers/ai2/): A US non-profit research institute. Its Olmo models are "fully open": it publishes the weights together with the training data, code and intermediate checkpoints, so that researchers can reproduce and study how a model is built. - [Aleph Alpha](https://ai-opensource.com/developers/aleph-alpha/): A German AI company that builds models for governments and regulated industries and stresses "sovereign" supply chains: training in Europe under European law. Its Kolibri 1 model is released under Apache 2.0. - [Alibaba Qwen](https://ai-opensource.com/developers/alibaba-qwen/): Alibaba's Qwen team. It publishes one of the widest open-weight model families, from small dense models to a 2.4-trillion-parameter flagship, with licenses that range from Apache 2.0 to custom Qwen terms. - [Alibaba Wan](https://ai-opensource.com/developers/alibaba-wan/): Alibaba's video-generation team. Its Wan 2.1 and Wan 2.2 models, and the character-animation models built on them, are open weights under Apache 2.0; the newer Wan versions are offered only through Alibaba's apps and API. - [AutoTrust AI Lab](https://ai-opensource.com/developers/autotrust/): A lab that publishes the Decider decision models on Hugging Face under Apache 2.0, built on Qwen and Gemma backbones and also available on ModelScope, plus a news recommendation and search demo at news.guru.so. We found little public information about the organization beyond its Hugging Face and ModelScope profiles, and its model cards still use the older JEV and GEV names. - [BAAI](https://ai-opensource.com/developers/baai/): The Beijing Academy of Artificial Intelligence, a Chinese research institute. Its BGE embedding and reranker models, published through the FlagOpen GitHub organization, are among the most downloaded retrieval models on Hugging Face. Its AREX-2 agent model, released in September 2026 under Apache 2.0, is built on Qwen3.8-27B. - [Baidu](https://ai-opensource.com/developers/baidu/): Chinese internet and AI company behind the ERNIE models and the PaddlePaddle deep-learning framework. Its open releases include compact document-parsing models under the MIT and Apache 2.0 licenses. - [Bilibili](https://ai-opensource.com/developers/bilibili/): A Chinese video platform. Its Index LLM team publishes open-weight models on Hugging Face as IndexTeam, led by the Index-Translate family, which covers text, image, speech and long-document translation. - [Black Forest Labs](https://ai-opensource.com/developers/black-forest-labs/): A German AI company, founded by researchers behind Stable Diffusion, that builds the FLUX image models. It sells API access to its largest models and publishes smaller ones as open weights under its own FLUX license. - [Boson AI](https://ai-opensource.com/developers/boson-ai/): Boson AI USA, Inc. builds the Higgs audio models for voice AI and sells them through a hosted API. It also publishes weights on Hugging Face, including Higgs TTS 3, under its own research and non-commercial license with a free grant for monetized creator content. - [BreezeBlue](https://ai-opensource.com/developers/breezeblue/): BreezeBlue is the trade name of Resonia, Inc., which runs a hosted voice platform and API. It publishes the Breeze TTS 2 text-to-speech weights for research and non-commercial use, with Apache 2.0 inference code and commercial terms only through its own service. - [Bria AI](https://ai-opensource.com/developers/bria/): A Tel Aviv company that sells visual generative AI trained on licensed data, with a hosted API and commercial licenses. It publishes non-commercial open weights under the FIBO name, led by the FIBO 2.0 image model. - [Cactus Compute](https://ai-opensource.com/developers/cactus-compute/): A San Francisco startup from Y Combinator's summer 2025 batch. It builds tiny on-device AI models and the open-source Cactus inference engine for phones, wearables, robots and other low-power hardware. Its Needle, Whistle and Mirage models are released under Apache 2.0. - [CircleStone Labs](https://ai-opensource.com/developers/circlestone-labs/): CircleStone Labs LLC publishes Anima, a 2B text-to-image model for anime and illustration that it made together with Comfy Org. The weights use its own non-commercial license, while the images you generate may be used commercially. - [Cloudflare](https://ai-opensource.com/developers/cloudflare/): A US network and cloud company that runs Workers AI, a hosted service for AI models. Clef, a set of decision models post-trained from Qwen and released under Apache 2.0, is the first model its Workers AI team trained. - [Cohere](https://ai-opensource.com/developers/cohere/): A Canadian enterprise-AI company whose research arm, Cohere Labs, publishes open weights. Its earlier Command models were released for non-commercial use under CC BY-NC; Command A+ moved to Apache 2.0. - [Datalab](https://ai-opensource.com/developers/datalab/): US company that builds the Marker, Surya and Chandra document-conversion tools. It publishes its OCR models openly, under a modified Open RAIL-M license that is free for research, personal use and small organizations. The limit is US$5 million of revenue and funding for Surya OCR 2 and US$2 million for Chandra OCR 2. - [Deep Principle](https://ai-opensource.com/developers/deep-principle/): Deep Principle (Chinese: 深度原理) builds AI for materials and chemistry research, including Mira, an AI scientist that connects models to simulations and laboratory experiments. It published Hydrogen-1.0, a science-focused post-train of GLM-5.3, under the MIT license. - [DeepReinforce](https://ai-opensource.com/developers/deepreinforce/): California AI research startup that publishes the Ornith agentic-coding models under the ornith-ai organization on Hugging Face. Ornith 1.0 and 1.5 are post-trained from Qwen3.5 and Gemma 4 and use the MIT license, with no regional limits. - [DeepSeek](https://ai-opensource.com/developers/deepseek/): Chinese AI lab known for efficient, large mixture-of-experts models. It releases its weights under the MIT license and publishes detailed technical reports. - [Fish Audio](https://ai-opensource.com/developers/fish-audio/): Voice AI company known for the open-source Fish Speech project. The legal entity named in its license is 39 AI, Inc., with offices in Palo Alto and Tokyo. Drama 3, released in October 2026, is a 36B mixture-of-experts text-to-speech model under the Fish Audio Community License, while the smaller S2 Pro from March 2026 stays under a research and non-commercial license. - [Google DeepMind](https://ai-opensource.com/developers/google-deepmind/): Google's AI research lab. Its open-weight Gemma family spans phone-sized to workstation-class models and, since Gemma 4, is released under the Apache 2.0 license. Its Magenta team also publishes Magenta RealTime 2, an open music model that plays live on a MacBook. - [Gradium](https://ai-opensource.com/developers/gradium/): A Paris voice-AI company, spun out of the Kyutai lab and founded by its cofounders, with offices in Paris and San Francisco. It sells voice APIs and open-sources its on-device Phonon text-to-speech models. The English Phonon 31m uses Apache 2.0, while the larger 40m models are CC BY-NC. - [HeartMuLa](https://ai-opensource.com/developers/heartmula/): The HeartMuLa team publishes open music foundation models: the HeartMuLa song generator, the HeartCodec audio codec and HeartTranscriptor for lyrics recognition. These are Apache 2.0, while its September 2026 MuLaCover cover-song model is CC BY-NC 4.0. - [hexgrad](https://ai-opensource.com/developers/hexgrad/): An independent developer, known as hexgrad on Hugging Face and GitHub and as rzvzn on Discord, who trained and released the Kokoro text-to-speech models. Kokoro-82M is an 82-million-parameter model under Apache 2.0 and one of the most downloaded open speech models. We found no published country or employer for the developer, so none is shown. - [HiDream.ai](https://ai-opensource.com/developers/hidream/): An image-generation company that publishes its models on Hugging Face under the HiDream-ai organization, from the HiDream-I1 and HiDream-E1 models of 2025 to HiDream-O1-Image, an 8B model released under the MIT license in May 2026. - [IBM](https://ai-opensource.com/developers/ibm/): The technology company behind the Granite family of open models for enterprise use, covering language, speech, embedding and guardrail models. Most Granite models use Apache 2.0, but at least one speech variant is non-commercial, so check each repository. - [Ideogram](https://ai-opensource.com/developers/ideogram/): The company behind the Ideogram image-generation service and its hosted API. Ideogram 4, published in June 2026, is its first open-weight model, with non-commercial weights and Apache 2.0 inference code. - [Institute of Foundation Models (MBZUAI)](https://ai-opensource.com/developers/ifm/): The Institute of Foundation Models (IFM) is a research institute at Mohamed bin Zayed University of Artificial Intelligence (MBZUAI) in Abu Dhabi, and the team behind LLM360. It publishes the K2 Horizon models and Jais 3, with weights and code under Apache 2.0, plus intermediate checkpoints, training logs and data or data recipes. - [inclusionAI (Ant Group)](https://ai-opensource.com/developers/inclusion-ai/): The open AI initiative founded by Ant Group, the Chinese fintech company. It publishes the Ling language models and related research, plus image models such as Ming-Image 0.1 Design; the recent Ling-3.1 and Ming-Image releases use the MIT license. - [JetBrains](https://ai-opensource.com/developers/jetbrains/): JetBrains is a software company, founded in Prague in 2000 and headquartered in Amsterdam, that makes developer tools such as IntelliJ IDEA and PyCharm. Its Mellum models, built for code and coding agents, are published with open weights under Apache 2.0. - [Jina AI](https://ai-opensource.com/developers/jina-ai/): A Berlin-based search-AI company that makes multilingual and multimodal embedding models, rerankers and small language models. Elastic completed its acquisition of Jina AI in October 2025, and the latest v5 models are still published on Hugging Face, under a non-commercial license. - [Kandinsky Lab](https://ai-opensource.com/developers/kandinsky-lab/): The research team behind the Kandinsky image and video diffusion models, which it publishes with code on GitHub and Hugging Face. Kandinsky 6.0 Video, released in October 2026 under the MIT license, generates five-second clips with synchronized audio. - [KittenML (Stellon Labs)](https://ai-opensource.com/developers/kittenml/): Voice AI company that publishes the Kitten-TTS models under the KittenML name, with Stellon Labs Inc. named as the copyright holder in its license. Its older Kitten-TTS 0.8 models run on CPU from 15M to 80M parameters, while Kitten-TTS 2 is a 1.7B ternary model under the Stellon Labs Community License, which is free until you pass US$1 million in revenue or funding. - [KRAFTON](https://ai-opensource.com/developers/krafton/): KRAFTON is a South Korean video-game company, and KRAFTON AI is its AI team. In October 2026 it released Raon-1 mini preview, a 28.8B mixture-of-experts language model trained from scratch under Apache 2.0, with strong Korean ability. - [Krea](https://ai-opensource.com/developers/krea/): Krea.ai, Inc., a Delaware corporation that runs a creative app for images, video and 3D. Its open release is Krea 2, an image model trained from scratch in a Raw base and a Turbo version. Its newer Krea 3 is not on Hugging Face. - [Kyutai](https://ai-opensource.com/developers/kyutai/): A non-profit open-science AI lab in Paris, founded in November 2023 by the Iliad Group, CMA CGM and Schmidt Sciences. It is best known for the Moshi speech dialogue model and the Mimi audio codec. Its Pocket TTS runs a voice-cloning speech model on a laptop CPU; the code is MIT and the weights are CC BY 4.0. - [LG AI Research](https://ai-opensource.com/developers/lg-ai-research/): The AI research arm of South Korea's LG Group, behind the EXAONE and K-EXAONE model families. K-EXAONE 2.0, released on 31 July 2026 for the government-backed sovereign AI foundation model project, is a 750B-parameter mixture of experts under Apache 2.0. - [LightOn](https://ai-opensource.com/developers/lighton/): French enterprise AI company that sells an AI search platform for company documents. Its open releases include the compact LightOnOCR document models under the Apache 2.0 license. - [Liquid AI](https://ai-opensource.com/developers/liquid-ai/): A US company building small, fast "Liquid Foundation Models" for phones, laptops and other edge devices, using an architecture that mixes short convolutions with attention. Besides the LFM language models it publishes small decision models such as Liquid d1. Its models use the LFM Open License, which is free commercially below US$10 million of annual revenue. - [LTX (Lightricks)](https://ai-opensource.com/developers/ltx/): An Israeli company behind the LTX video models. LTX became a separate business unit of Lightricks, the maker of consumer creative apps, in June 2026; its Hugging Face and GitHub organizations still use the Lightricks name. It publishes open weights free for organizations under US$10 million in annual revenue. LTX-2.3 still uses the January 2026 LTX-2 license, while LTX-2.5 uses the newer LTX-2.x license. - [M-A-P](https://ai-opensource.com/developers/m-a-p/): M-A-P (Multimodal Art Projection) is an open-source research community behind open music models such as YuE, MERT and SheetSage. Its YuE2 song generator, released in September 2026 with Tokenwave.AI, MBZUAI and ACE Studio, has weights under CC BY-NC 4.0, so it is non-commercial only. - [Maritaca AI](https://ai-opensource.com/developers/maritaca-ai/): A Brazilian company that builds language models for Brazilian Portuguese and offers them through its own chat and API. Its flagship S5 is served that way, while S5-nano, released in November 2026 on top of Qwen3.5-4B under Apache 2.0, is open-weight. - [Maya Research](https://ai-opensource.com/developers/maya-research/): A voice-AI startup, backed by a US$1.9 million seed round from South Park Commons, that builds speech models for Indian languages and sells them through a hosted API. It publishes the Maya and Veena text-to-speech models on Hugging Face, and Maya 2 Native uses Apache 2.0. - [Meituan](https://ai-opensource.com/developers/meituan/): Chinese food-delivery and local-services platform. Its LongCat team releases open-weight language models, including LongCat-2.0 in July 2026 and the 1.6T-parameter multimodal LongCat-2.5 in November 2026, which uses the MIT license. - [MemTensor](https://ai-opensource.com/developers/memtensor/): A Chinese company behind MemOS, an open-source memory framework for LLM agents. Its MemT-1.0 models build a memory state into the model itself, with Nano weights under Apache 2.0 and Flash weights under MIT. - [Meta](https://ai-opensource.com/developers/meta/): The company behind the Llama models. Its newer Muse series includes Muse Glimmer, a small Apache 2.0 model for local use, and the much larger Muse Spark Open under a custom license. - [Microsoft](https://ai-opensource.com/developers/microsoft/): The US technology company. Its research teams publish open models on Hugging Face and GitHub, often under the MIT license, including the VibeVoice speech models from Microsoft Research and the Harrier text embedding models. Its Mage-VL vision-language models use Apache 2.0. - [MiniMax](https://ai-opensource.com/developers/minimax/): Shanghai-based AI company behind the MiniMax-M series of efficient mixture-of-experts language models, the H3 video and image models and the Music 3 song generator. Its licenses differ by model. M2.5 uses a permissive-style license with notice and prohibited-use rules, M2.7 is limited to non-commercial use, and the M3 community license allows free non-commercial use and sets notice, credit and revenue rules for commercial use. The H3 licenses also exclude the US, EU, UK and South Korea, and the Music 3 license asks for an interface credit and written authorization above US$20 million in revenue. - [Mistral AI](https://ai-opensource.com/developers/mistral-ai/): Paris-based AI company and Europe's best-known open-weight lab. It positions its models for sovereign, self-hosted AI; Mistral Large 4 uses a modified MIT license with a rule for hosting businesses. - [Moonshot AI](https://ai-opensource.com/developers/moonshot-ai/): Chinese AI company behind the Kimi assistant. Its Kimi K3 and K3.1 Pro models are among the largest open-weight models, released under a custom license with conditions for large hosting businesses. The earlier K2.5, K2.6 and K2.7 Code models use a Modified MIT license that only asks very large products to show the model name. - [Motif Technologies](https://ai-opensource.com/developers/motif-technologies/): South Korean AI company that designs its models, tokenizer and training stack in-house. Motif 3, a 314B-parameter mixture of experts released in August 2026 under the MIT license, was built for the government-backed sovereign AI foundation model project. - [Nanbeige LLM Lab](https://ai-opensource.com/developers/nanbeige/): Nanbeige (Chinese: 南北阁) is the large-language-model research team at BOSS Zhipin, a Chinese recruitment platform. It publishes open-weight agent models under the Nanbeige4 name, from 3B-class models to the 33B Nanbeige4.3, and its model cards list Apache 2.0 without a LICENSE file. - [NetEase Youdao](https://ai-opensource.com/developers/netease-youdao/): Chinese company with offices in Beijing that publishes the Confucius model family on Hugging Face, including Confucius4, a text-to-speech model and the streaming speech recognizer Confucius4-R2T2, and the QAnything question-answering project. R2T2 uses the NetEase Youdao Model Use License Agreement, which is free commercially below 100 million monthly users and RMB 1 billion of annual revenue. - [Nex-AGI](https://ai-opensource.com/developers/nex-agi/): An open-source foundation-model group whose GitHub organization, SII Nex, says it is built by Shanghai Innovation Institute and partners. Its Nex-N2 agent models build on open bases such as DeepSeek-V4-Pro-Base and GLM-5.3-Flash and are tagged Apache 2.0. - [Nous Research](https://ai-opensource.com/developers/nous-research/): A US AI research company that trains open models, including the Hermes line, and builds Hermes Agent, an open-source agent harness. Its Arete-27B, released in November 2026, is a post-training of Alibaba's Qwen3.8-27B under Apache 2.0. - [NuMind](https://ai-opensource.com/developers/numind/): A document-AI company that builds specialized models for structured data extraction and OCR and sells them through the NuExtract platform. It publishes its models on Hugging Face, and says its NuExtract line has passed 2 million downloads and its NuMarkdown-Thinking-8B 1.5 million. NuExtract3 uses Apache 2.0. - [NVIDIA](https://ai-opensource.com/developers/nvidia/): The GPU maker, which also publishes open models, datasets and training recipes under the Nemotron name, plus speech, robotics and biology models. Recent Nemotron releases use the permissive OpenMDW-1.1 license, Nemotron 3 Nano Omni uses the NVIDIA Open Model Agreement, and LocateAnything-3B has a research-only NVIDIA License. - [OpenAI](https://ai-opensource.com/developers/openai/): The company behind ChatGPT. Its gpt-oss models, released in August 2025, are its open-weight reasoning models under Apache 2.0. Its Privacy Filter, released in April 2026 under Apache 2.0, is a small model that finds personal data in text, and it also publishes the Whisper speech-recognition models. - [OpenBMB](https://ai-opensource.com/developers/openbmb/): A Chinese open-source model community that releases the MiniCPM, AgentCPM and VoxCPM families together with the company ModelBest. It focuses on small, efficient models for phones and laptops, and its recent releases use Apache 2.0. - [OpenMOSS](https://ai-opensource.com/developers/openmoss/): OpenMOSS is a research group under the Shanghai Innovation Institute that works closely with Fudan University and MOSS-AI. It publishes open models under the MOSS name. They include the MOSS-TTS speech family, built with MOSI.AI, with MOSS-TTS, MOSS-TTS-Nano and MOSS-TTSD under Apache 2.0, MOSS-SoundEffect v2.0, a 1.3B text-to-sound-effect model under Apache 2.0, MOSS-Image, a 5B text-to-image model released in October 2026 under Apache 2.0, and MOSS-Transcribe-Diarize, a 0.9B model that transcribes long multi-speaker recordings with speaker labels. - [Perplexity](https://ai-opensource.com/developers/perplexity/): The US company behind the Perplexity answer engine. On Hugging Face it publishes the pplx-embed embedding models, most of them under the MIT license, and the pplx-decider decision models under Apache 2.0, along with other models such as PII-Tracer and BrowseSafe. - [Phạm Nguyễn Ngọc Bảo (VieNeu-TTS)](https://ai-opensource.com/developers/pham-nguyen-ngoc-bao/): Independent developer who builds the VieNeu-TTS project for Vietnamese speech, published on Hugging Face under the pnnbao-ump account and on GitHub as pnnbao97. The v3 Turbo model card says its architecture is the author's own design and that it was trained from scratch, and the project ships a Python SDK and an OpenAI-compatible speech server. The weights are under Apache 2.0. - [Pika](https://ai-opensource.com/developers/pika/): Pika is a US video-generation company with hosted video apps and an API. In October 2026 it published Pika H3 Flex, a post-trained version of MiniMax H3 that makes video with sound in six sampling steps, as open adapters under the MiniMax H3 Community License. - [Poolside](https://ai-opensource.com/developers/poolside/): San Francisco AI company, founded in 2023, that builds models for agentic coding. Its Laguna series began with the Apache 2.0 Laguna XS.2 in April 2026, Poolside says open weights are now its default, and Laguna XS 2.1 and S 2.1 use the OpenMDW-1.1 license. - [Preferred Networks](https://ai-opensource.com/developers/preferred-networks/): Preferred Networks (PFN) is a Japanese AI company that develops the PLaMo language models and its own MN-Core AI accelerators. It publishes PLaMo 3 Translate and other PLaMo models under the PLaMo Community License, which is free for personal and academic use and, after registration, for commercial users below a revenue limit. - [PrismML](https://ai-opensource.com/developers/prismml/): A company based in Pasadena, California, also written Prism ML, that builds very low-bit AI models. Its Bonsai models compress open base models such as Qwen to ternary or 1-bit weights, so that large models fit on laptops, phones and single GPUs. The Bonsai 2 models are released under Apache 2.0. - [Pruna AI](https://ai-opensource.com/developers/pruna-ai/): Pruna AI is a company founded in 2023 and based in Munich, with a second research hub in Paris. It is known for its open-source library that compresses and speeds up AI models, and it now also publishes its own efficient image and video models, some of them reworked from other developers' open weights. - [Recursive](https://ai-opensource.com/developers/recursive/): Recursive Superintelligence, Inc. says its goal is recursively self-improving superintelligence that automates knowledge discovery, and it builds its models with an automated research system it calls the Eureka Machine. Flint, a 40B coding model released in October 2026 under Apache 2.0, is its first open-weight model. It is a different company from Ricursive Intelligence. - [Reflection AI](https://ai-opensource.com/developers/reflection-ai/): US AI lab that released its first open-weight model, the 501-billion-parameter Beam, under Apache 2.0 in October 2026. - [Resemble AI](https://ai-opensource.com/developers/resemble-ai/): California voice-AI company that open-sourced the Chatterbox and DramaBox voice models under the MIT license. In August 2026 it said it was no longer selling voice generation to new customers and would focus on deepfake detection, leaving its voice models as open, unmaintained research. - [Ricursive Intelligence](https://ai-opensource.com/developers/ricursive-intelligence/): US AI lab that builds self-improving systems by co-designing hardware, software and models. A proprietary model powers its AI chip-design platform, and in November 2026 it released a general-purpose variant, the 65-billion-parameter Ripple-65B Preview, under Apache 2.0. - [Sarvam AI](https://ai-opensource.com/developers/sarvam-ai/): An Indian company that builds AI for Indian languages and enterprises. Its Sarvam 30B and 105B reasoning models, released in March 2026 and trained from scratch in India on IndiaAI mission compute, and its Sarvam Prism decision model use Apache 2.0. - [Sber (GigaChat)](https://ai-opensource.com/developers/sber/): The GigaChat team at Sber publishes open weights under the ai-sage name on Hugging Face and on GitVerse, and announces each release on Sber's corporate Habr blog. GigaChat 3.5 Ultra and its Reasoning version are 432B mixture of experts models, GigaChat 3.1 Lightning is a 10B model with 1.8B active parameters, and the GigaAM family covers speech recognition for Russian and Central Asian languages. All of them are under the MIT license. - [SenseTime](https://ai-opensource.com/developers/sensetime/): Chinese AI company whose SenseNova team publishes open models under the OpenSenseNova GitHub organization. SenseNova-U1.5 (August 2026) is a native unified image-and-text model, and SenseNova-Loop-9B-Preview (30 October 2026) is a 9B looped vision-language model, both under Apache 2.0. - [Shanghai AI Laboratory](https://ai-opensource.com/developers/shanghai-ai-laboratory/): A Chinese AI research laboratory in Shanghai. Its AI for Science Center runs the InternScience open-source organization, which publishes the Agents-A1 and Sage-A1.5 agent models under Apache 2.0. - [SK Telecom](https://ai-opensource.com/developers/sk-telecom/): South Korean telecommunications company behind the A.X family of language models. A.X K2, released in July 2026, is a 688B-parameter mixture of experts under Apache 2.0, built for the government-backed sovereign AI foundation model project. - [Stability AI](https://ai-opensource.com/developers/stability-ai/): Stability AI is the UK company behind Stable Diffusion. Its Stable Audio 3 family, released in May 2026, is its newest open-weight line: the Small and Medium models were trained on licensed and Creative Commons data and use the Stability AI Community License, which is free for commercial use below US$1 million in annual revenue. - [StepFun](https://ai-opensource.com/developers/stepfun/): StepFun (阶跃星辰) is a Chinese AI company behind the Step family of language models, with Step-Audio speech models and Step1X image-editing models alongside them. It publishes weights on Hugging Face and ModelScope, and its newest flagship, Step 5 Preview, comes with a custom license. - [Supertone](https://ai-opensource.com/developers/supertone/): A South Korean voice-AI company that built the on-device Supertonic text-to-speech models, which run through ONNX Runtime. Supertonic 3 covers 31 languages, with the weights under the BigScience OpenRAIL-M license. In September 2026 the company archived its open-source repositories and said development and support had ended. - [Swiss AI Initiative](https://ai-opensource.com/developers/swiss-ai/): A Swiss research initiative led by ETH Zurich, EPFL and the Swiss National Supercomputing Centre (CSCS). Its Apertus models are fully open: weights, data preparation, training code and evaluations are published under permissive licenses, and training uses only openly available data. - [Tencent](https://ai-opensource.com/developers/tencent/): Chinese internet company behind WeChat and the Hunyuan (Hy) model family. Its Hy3 and Hy4 preview language models, the Hy-MT2 translation models and the Hy-Nomas interactive video world model use Apache 2.0, while HunyuanOCR still uses a Hunyuan community license that excludes the EU, the UK and South Korea. Its AuK speech generation and editing model uses the MIT license, so check each repository. - [Thinking Machines Lab](https://ai-opensource.com/developers/thinking-machines-lab/): US AI lab that builds tools for customizing models. Its first open-weight family, Inkling, is released under Apache 2.0 and is designed as a base for fine-tuning. - [TokenSpark (SparkLLM)](https://ai-opensource.com/developers/tokenspark/): Company behind the Spark model family, which publishes on Hugging Face as XHToken. Its profile says it develops general-purpose foundation models and professional AI agent products, including two harness systems called AStudio and Loomy. The Spark-X2.5 and Spark-X2.6 on-device language models in 1.7B and 4B sizes use Apache 2.0. - [Tongyi MAI](https://ai-opensource.com/developers/tongyi-mai/): A team in Alibaba Group's Tongyi Lab that builds efficient generative models. It is best known for the 6B Z-Image image models, released under Apache 2.0, and also publishes MAI-UI, a family of GUI agent models. - [Tongyi Lab Speech Team](https://ai-opensource.com/developers/tongyi-speech/): The speech team of Alibaba Group's Tongyi Lab, which publishes its work under the FunAudioLLM name. It is the team behind the CosyVoice speech-synthesis models, released under Apache 2.0. - [Upstage](https://ai-opensource.com/developers/upstage/): South Korean AI company behind the Solar family of language models. Solar Open 2, released in July 2026, is a 250B-parameter mixture of experts under the Upstage Solar License, which is Apache 2.0 plus a naming rule for derived models. - [Video Rebirth](https://ai-opensource.com/developers/video-rebirth/): Video Rebirth is a Singapore-based AI video start-up founded in 2024. Besides its hosted video models, it publishes the HyperFlow series, which compresses open video models such as MiniMax H3 so they need fewer sampling steps and less memory. - [Voyage AI](https://ai-opensource.com/developers/voyage-ai/): Voyage AI builds the Voyage text embedding models, which its blog says are available through the Voyage API and, for MongoDB Atlas customers, through the Atlas Embedding and Reranking API. voyage-4-nano, released with the Voyage 4 series on 15 January 2026, is its first open-weight model under Apache 2.0, and its notice file carries a MongoDB, Inc. copyright. - [webAI](https://ai-opensource.com/developers/webai/): A company that makes the webAI app, a workspace where AI models run locally on the user's device. Its TwIL reasoning models, post-trained from IBM Granite 4.2 and SmolLM3, use the webAI Non-Commercial License rather than the Apache 2.0 license of their base models. - [Xiaomi](https://ai-opensource.com/developers/xiaomi/): Chinese consumer-electronics maker. Its MiMo foundation-model team publishes large agent-focused models and small research checkpoints under MIT license tags on Hugging Face, and other Xiaomi research teams release models such as Xiaomi-OCR-0 and the OmniVoice speech model. - [Z.ai (Zhipu AI)](https://ai-opensource.com/developers/z-ai/): Chinese lab developing the GLM model series. GLM-5.3-Flash is MIT licensed, while the larger GLM-5.3 uses a custom license that only restricts very large hosting businesses. GLM-5, 5.1, 5.2, the 30B-A3B GLM-4.7-Flash and the GLM-OCR model are all MIT licensed too. - [Zyphra](https://ai-opensource.com/developers/zyphra/): San Francisco AI lab behind the Zonos text-to-speech line. ZONOS3-Mini, released in November 2026 under Apache 2.0, is an 800M-parameter mixture-of-experts voice-cloning model that Zyphra says runs in real time on laptops, phones and in a browser tab. ## News - [Apertus 1.5: Switzerland's fully open model reads images and audio](https://ai-opensource.com/news/apertus-1-5-released/): 2026-11-11. ETH Zurich, EPFL and CSCS released Apertus 1.5 on 24 July: fully open 8B and 70B models with image and audio input, a 262K context and Apache 2.0. Audio understanding is experimental. - [Clef: Cloudflare's open decision models for agent workflows](https://ai-opensource.com/news/clef-released/): 2026-11-11. Cloudflare released Clef and Clef-flash under Apache 2.0: 27B and 9B decision models, post-trained from Qwen, that return a probability for each option you define instead of writing text. - [Cosmos 3: NVIDIA's open world models for robots and driving](https://ai-opensource.com/news/cosmos-3-released/): 2026-11-11. NVIDIA released Cosmos 3 under OpenMDW-1.1: 4B, 16B and 64B world models that generate video, images and robot actions and reason about video, for robotics and autonomous driving. - [EmbeddingGemma 2 embeds text, images, video and audio under Apache 2.0](https://ai-opensource.com/news/embeddinggemma-2-released/): 2026-11-11. Google DeepMind's 740M EmbeddingGemma 2 maps text, code, images, video and audio into one 768-dimension space, runs on-device and moves from the Gemma license to Apache 2.0. - [Baidu's ERNIE 5.5 Flash Base: 280B parameters, license excludes the EU](https://ai-opensource.com/news/ernie-5-5-flash-base-released/): 2026-11-11. Baidu released ERNIE-5.5-Flash-Base on 5 November: a 280B mixture-of-experts base model with 14B active and a 1M context. Its license excludes the EU, the UK and South Korea. - [Four open translation models, four different yardsticks](https://ai-opensource.com/news/four-open-translation-models-four-yardsticks/): 2026-11-11. TranslateGemma, Hy-MT2, Index-Translate and PLaMo 3 Translate each report quality on different test sets, metrics and rivals, so their claims cannot be compared directly. - [Free until you grow: where open licenses draw the line](https://ai-opensource.com/news/free-until-you-grow-open-license-thresholds/): 2026-11-11. Thirteen licenses in our catalog end free commercial use at a revenue or user threshold, from US$500,000 to US$50 million. Here is how the lines, the counting and the duties differ. - [Tencent's Hy4 preview: 770B parameters, 1M context and Apache 2.0](https://ai-opensource.com/news/hy4-preview-released/): 2026-11-11. Tencent's Hy4 preview is a 770B mixture-of-experts model with 49B active and a 1M-token context, under Apache 2.0 after Hy3's preview had territory limits. Tencent calls it an early version. - [Hydrogen-1.0: a 753B open science model built on GLM-5.3](https://ai-opensource.com/news/hydrogen-1-0-released/): 2026-11-11. Deep Principle released Hydrogen-1.0, a 753B science model post-trained from GLM-5.3 under an MIT label, with 1M context and 4-bit builds. All benchmark claims are its own. - [Index-Translate: Bilibili's open translators for 150 languages](https://ai-opensource.com/news/index-translate-released/): 2026-11-11. Bilibili's Index team released Index-Translate, Apache 2.0 translation models in 2B, 9B and 35B-A3B sizes for 150 languages, with image translation and glossary, format and tone controls. - [Korea's four sovereign AI models: open weights, three teams advance](https://ai-opensource.com/news/korea-sovereign-ai-four-open-models/): 2026-11-11. Four open-weight models of 250B parameters or more came out of South Korea's sovereign AI project. Compare their sizes, licenses and hardware needs, and see which teams advanced. - [Krea 2: an open image model trained from scratch for variety](https://ai-opensource.com/news/krea-2-released/): 2026-11-11. Krea released Krea 2, a 12.8B text-to-image model trained from scratch, as a Raw base and an 8-step Turbo. Commercial use is free only below US$1 million in annual revenue. - [Meituan's LongCat-2.5: 1.6T parameters, vision and the plain MIT license](https://ai-opensource.com/news/longcat-2-5-released/): 2026-11-11. Meituan released LongCat-2.5 as open weights on 9 November: a 1.6T mixture of experts with 48B active, image and video input, a 1M-token context claim and the MIT license. - [MiniMax H3-Image: 16-megapixel open weights, but not for the US or EU](https://ai-opensource.com/news/minimax-h3-image-released/): 2026-11-11. MiniMax's 33B H3-Image makes 16-megapixel pictures in six steps and edits with up to 16 references. Its license excludes the US, EU, UK and South Korea and caps commercial use at US$1M. - [MOSS-Image: a 5B open image model for long text and JSON layouts](https://ai-opensource.com/news/moss-image-released/): 2026-11-11. OpenMOSS released MOSS-Image under Apache 2.0: a 5B generator that renders long English and Chinese text and follows JSON layouts, with a low-memory profile near 4 GB of VRAM. - [Nanbeige4.3-33B: a 33B agent model with a 23 GB 4-bit build](https://ai-opensource.com/news/nanbeige-4-3-33b-released/): 2026-11-11. BOSS Zhipin's Nanbeige lab published Nanbeige4.3-33B on 24 October: a dense 33B text model with a 512K context, an INT4 build of 22.6 GB and an Apache 2.0 tag but no LICENSE file. - [NVIDIA's Nemotron 3.5 Embed and Parse: document pages as images](https://ai-opensource.com/news/nemotron-3-5-embed-and-parse-released/): 2026-11-11. NVIDIA released Nemotron 3.5 Embed, which searches page images and text without OCR, and Nemotron 3.5 Parse, a 631M page parser, both under OpenMDW-1.1 for commercial use. - [Open models that listen: nine take audio, but only three answer in speech](https://ai-opensource.com/news/open-models-that-listen-nine-take-audio-three-speak/): 2026-11-11. Gemma 4, Apertus 1.5, Inkling, MiMo-V2.6, Muse Spark Open, Nemotron 3 Nano Omni, Qwen3-Omni, MiniCPM-o 4.5 and LFM3-Omni compared by audio input, speech output, size and license. - [Open music models grew up in 2026: seven models for songs, instruments and sound effects](https://ai-opensource.com/news/open-music-models-seven-models-four-jobs/): 2026-11-11. Seven open music and sound models arrived between January and September 2026, from song writers such as ACE-Step 1.5 to a live instrument from Google. Compare jobs, hardware and licenses. - [Eleven open OCR models in a year: small and hard to compare](https://ai-opensource.com/news/open-ocr-models-small-and-hard-to-compare/): 2026-11-11. Eleven open OCR models arrived between October 2025 and October 2026. Most have 1B to 5B parameters and Apache 2.0 or MIT licenses, but benchmark versions differ, so their scores cannot be ranked. - [Open speech models: transcription is open, text-to-speech is mixed](https://ai-opensource.com/news/open-speech-models-transcription-open-voices-mixed/): 2026-11-11. Ten of our eleven open speech-to-text models have no commercial limit, but only six of fourteen text-to-speech models are plain Apache 2.0. The rest are non-commercial, capped, restricted or retired. - [PLaMo 3 Translate: PFN's open translators for 53 languages](https://ai-opensource.com/news/plamo-3-translate-released/): 2026-11-11. Preferred Networks released PLaMo 3 Translate in 0.8B to 31B sizes for 53 languages, with GGUF files from 0.57 GB and a license that limits commercial use to 1 billion yen of revenue. - [Raon-1: KRAFTON's Korean-first 28.8B model and a 3.1B on-device sibling](https://ai-opensource.com/news/raon-1-released/): 2026-11-11. KRAFTON AI released Raon-1 mini preview, a 28.8B mixture of experts trained from scratch, and the 3.1B Raon-1 nano distilled from it, both under Apache 2.0 with a 262K context. - [Ripple-65B Preview: a dense 65B agent model under Apache 2.0](https://ai-opensource.com/news/ripple-65b-preview-released/): 2026-11-11. Ricursive Intelligence, a chip-design lab, released Ripple-65B Preview on 4 November: a dense 65B text model for agentic coding with 1M context, Apache 2.0 and a 57 GB NVFP4 build. - [Sber's open models: a 432B chat model, speech and embeddings](https://ai-opensource.com/news/sber-opens-gigachat-speech-and-embedding-models/): 2026-11-11. Sber's GigaChat team has five MIT-licensed entries on Hugging Face: GigaChat 3.5 Ultra, 3.1 Lightning, GigaAM-v3, GigaAM Multilingual and Giga-Embeddings. Sizes, scores and caveats. - [PrismML's Ternary Bonsai 2: a 27B-class model in 5.95 GB](https://ai-opensource.com/news/ternary-bonsai-2-released/): 2026-11-11. PrismML released Ternary Bonsai 2: Qwen models stored with three-value weights, from a 27B model in 5.95 GB to a 2B in 0.43 GB, with image input, a 262K context and Apache 2.0. - [Tiny text-to-speech models that run on a CPU: four open options compared](https://ai-opensource.com/news/tiny-text-to-speech-models-run-on-a-cpu/): 2026-11-11. Kokoro, Pocket TTS, Phonon 31M and Supertonic 3 are open text-to-speech models of 100M parameters or fewer that run on a laptop CPU. Compare sizes, languages, voice cloning and licenses. - [ZONOS3-Mini: open TTS in 43 languages that runs in a browser](https://ai-opensource.com/news/zonos3-mini-released/): 2026-11-11. Zyphra released ZONOS3-Mini on 9 November: an 800M mixture-of-experts text-to-speech model with 100M active parameters, 43 languages, voice cloning and Apache 2.0 weights. - [Giant open models now ship in 4-bit: what it means for memory planning](https://ai-opensource.com/news/giant-open-models-ship-in-4-bit/): 2026-11-08. Eight models in our catalog ship their main checkpoint in a 4-bit-class format, and nineteen more add official 4-bit builds. See the file sizes and how they change the hardware you need. - [Open licenses: free to build on, but hosting needs a deal](https://ai-opensource.com/news/hosting-clauses-in-open-licenses/): 2026-11-08. Six of the biggest 2026 open-weights licenses let you build products freely but require a separate agreement or review from businesses that sell hosted access. Here is how the thresholds differ. - [Qwen3.8: three language models, three licenses, and a fourth for images](https://ai-opensource.com/news/qwen-3-8-three-models-three-licenses/): 2026-11-08. Alibaba's Qwen3.8 models ship under Apache 2.0, the Qwen3.8-Max License and Qwen Community 1.0, while Qwen-Image-2.1 is research-only. A map of what each license lets you do. - [Moonshot releases Kimi K3.1 Pro: 2.8T parameters, open weights](https://ai-opensource.com/news/kimi-k3-1-pro-released/): 2026-11-02. Moonshot AI published Kimi K3.1 Pro on 2 November 2026: a 2.8T mixture-of-experts model for long coding and office tasks, with a 1M-token context and 4-bit weights of about 1.56 TB. - [Mistral Large 4 goes open-weights: 1T parameters, trained in Europe](https://ai-opensource.com/news/mistral-large-4-open-weights/): 2026-10-29. Mistral published the weights of Mistral Large 4 on 29 October 2026: a multimodal mixture-of-experts model with a 1M-token context, 160+ languages and a license with a hosting condition. - [NVIDIA releases Nemotron 3.5 Transcribe, a promptable speech model](https://ai-opensource.com/news/nemotron-3-5-transcribe-released/): 2026-10-28. NVIDIA published Nemotron 3.5 Transcribe on 28 October 2026 under OpenMDW-1.1: 40 languages, speaker labels, translation and prompt-controlled transcript style, for recordings up to an hour. - [Liquid AI announces LFM3: six small models, text to voice](https://ai-opensource.com/news/liquid-ai-lfm3-family/): 2026-10-27. Liquid AI released the LFM3 family on 27 October 2026: text, vision, audio and omni models of about 2B parameters built for devices, plus LFM3-Audio-2B for voice, under the LFM Open License. - [MiniMax releases M3.1 Flash, a 428B model built for 4-bit](https://ai-opensource.com/news/minimax-m3-1-flash-released/): 2026-10-17. MiniMax published the M3.1 Flash weights on 17 October 2026: 428B parameters, 23B active, a 1M-token context and a 252 GB checkpoint, under a community license with credit and notice conditions. - [Black Forest Labs releases FLUX 3 Image Klein with layout control](https://ai-opensource.com/news/flux-3-image-klein-released/): 2026-10-15. Black Forest Labs published FLUX 3 Image Klein on 15 October 2026: 5.6B parameters, editing with up to 10 references, box-based layout control, an 8 GB INT8 build and a revenue-capped license. - [Meta releases Muse Spark Open, an 816B open-weights model](https://ai-opensource.com/news/muse-spark-open-released/): 2026-10-15. Meta published Muse Spark Open on 15 October 2026: 816B parameters, 42B active, text, image, video and audio input, a 1M-token context, a gated download and a 'Built with' credit requirement. ## Optional - [VRAM calculator](https://ai-opensource.com/vram-calculator/): Estimate how much GPU or unified memory an AI model needs at BF16, FP8 and 4-bit precision, and which hardware can run it. Pick a model or enter your own size. - [All models (JSON)](https://ai-opensource.com/data/models.json): Download the full catalog of open-source AI models as JSON or CSV: parameters, context window, license type, release date and sources. Free to reuse with attribution. - [All models (CSV)](https://ai-opensource.com/data/models.csv) - [Licenses (JSON)](https://ai-opensource.com/data/licenses.json) - [RSS feed](https://ai-opensource.com/rss.xml) - [llms-full.txt](https://ai-opensource.com/llms-full.txt): The full text of this site in one file