Best open-source LLMs in 2026

The open-weights language models worth knowing in November 2026, from trillion-parameter frontier models to a 27B model that fits one GPU. Each entry says what the model is for, how big it is, and what its license asks of you.

  • List updated November 11, 2026
  1. Kimi K3.1 Pro Largest

    The largest model here: 2.8T parameters with 104B active, a 1M-token context and image input, tuned for coding and agent work. It needs a multi-node cluster, and its MIT-style license adds conditions for large hosting businesses and very large products.

    • Moonshot AI
    • 2.8T (104B active)
    • 1M tokens
    • Kimi K3 License (Conditional)
    • from 1.72 TB
  2. Qwen3.8-Max

    Alibaba's first open Max-class model: 2.4T parameters, 95B active, 1M context, with text, image and video input. The MIT-style license is free for almost everyone, with a credit rule for huge products and a separate deal for large hosting businesses.

    • Alibaba Qwen
    • 2.45T (95B active)
    • 1M tokens
    • Qwen3.8-Max License (Conditional)
    • from 1.69 TB
  3. GLM-5.3

    A text-only 753B model with 40B active parameters, aimed at coding, with a 1M context. The license is close to MIT; only model-as-a-service operators with more than US$10 billion of revenue need a security review.

    • Z.ai (Zhipu AI)
    • 753B (40B active)
    • 1M tokens
    • GLM-5.3 License (Conditional)
    • from 520 GB
  4. LongCat-2.5 Newest

    Meituan's 1.6T multimodal mixture of experts with 48B active parameters, released on 9 November 2026 under the plain MIT license. Meituan documents deployment only on NPU servers, and the BF16 files total about 3.6 TB.

    • Meituan
    • 1.6T (48B active)
    • 1M tokens
    • MIT (Permissive)
    • from 1.1 TB
  5. DeepSeek V4.1 Flash

    A 552B mixture of experts with a tiny KV cache, native image input and a 1M context, under the MIT license. The small cache is what makes long-context serving cheaper.

    • DeepSeek
    • 552B (16B active)
    • 1M tokens
    • MIT (Permissive)
    • from 561 GB
  6. Mistral Large 4

    Europe's largest open-weight model: about 1T parameters with 52B active, multimodal, with a 1M context. It uses a modified MIT license, and businesses that sell hosted access need a commercial license from Mistral.

    • Mistral AI
    • 1.05T (52B active)
    • 1M tokens
    • Mistral Open-Weight (Conditional)
    • from 671 GB
  7. Muse Spark Open

    Meta's open-weight 816B model with 42B active parameters and a 1M context, taking text, image, video and audio. A business selling hosted access above US$20 million of revenue needs an agreement, everyone must display a credit line, and the download is gated.

    • Meta
    • 816B (42B active)
    • 1M tokens
    • Muse Model License (Conditional)
    • from 563 GB
  8. Inkling

    The largest Apache 2.0 model in our catalog: 975B multimodal mixture of experts from Thinking Machines, released as a base for customization rather than a finished assistant.

    • Thinking Machines Lab
    • 975B (41B active)
    • 1M tokens
    • Apache 2.0 (Permissive)
    • from 188 GB
  9. Reflection Beam

    Reflection's first open model: a 501B text mixture of experts under Apache 2.0 with 23.4B active parameters, built to be efficient at inference.

    • Reflection AI
    • 501B (23.4B active)
    • 1M tokens
    • Apache 2.0 (Permissive)
    • from 323 GB
  10. Ripple-65B Preview

    Ricursive Intelligence's Apache 2.0 65B dense preview with a 1M-token context, released on 4 November 2026. It is a text-only model aimed at long coding and agent tasks, and its BF16 weights are about 130 GB, so by our estimate it needs about 45 GB at 4-bit.

    • Ricursive Intelligence
    • 64.8B
    • 1M tokens
    • Apache 2.0 (Permissive)
    • from 45 GB
  11. Hydrogen-1.0

    Deep Principle's 753B science model, post-trained from GLM-5.3 on chemistry, materials and life-science data and released on 2 November 2026 under an MIT label. It has a 1M-token context, and its 4-bit builds are 423 to 464 GB. All benchmark claims are the developer's own.

    • Deep Principle
    • 753B (40B active)
    • 1M tokens
    • MIT (Permissive)
    • from 465 GB
  12. DeepSeek V4

    DeepSeek's MIT-licensed 1M-context pair: a 1.6T Pro with 49B active parameters and a 284B Flash with 13B active, built for long-context agents. By our estimate the Pro needs about 1 TB of memory at 4-bit and the Flash about 184 GB.

    • DeepSeek
    • 1.6T (49B active)
    • 1M tokens
    • MIT (Permissive)
    • from 176 GB
  13. Hy3

    Tencent's 295B mixture of experts with 21B active parameters and a 256K context. It moved to Apache 2.0 after a preview that came with a territory-limited license. By our estimate it needs about 200 GB at 4-bit.

    • Tencent
    • 295B (21B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 204 GB
  14. Laguna S 2.1

    Poolside's coding-focused 118B mixture of experts with 8B active parameters and a 1M context, under the permissive OpenMDW-1.1 license. By our estimate the 96 GB Q4_K_M file needs about 106 GB, so a 128 GB machine can run it. Its scores are Poolside's own.

    • Poolside
    • 118B (8B active)
    • 1M tokens
    • OpenMDW-1.1 (Permissive)
    • from 81 GB
  15. Step 5 Preview

    StepFun's 600B mixture of experts with 27B active parameters, a 1M context and image input, for coding agents, long tasks and finance. The FP8 weights are about 650 GB, and the custom license needs a separate deal from hardware makers and from hosting businesses above US$20 million a year.

    • StepFun
    • 604B (27B active)
    • 1M tokens
    • StepFun 1.0 (Conditional)
    • from 417 GB
  16. GigaChat 3.5 Ultra

    Sber's MIT-licensed 432B mixture of experts with 28B active and a 262K context, in Instruct and Reasoning versions. Sber's own tables put the Reasoning model level with DeepSeek V4 Flash Preview on AIME but well behind on SWE-bench Verified, and the FP8 weights alone are about 440 GB.

    • Sber (GigaChat)
    • 438B (28B active)
    • 256K tokens
    • MIT (Permissive)
    • from 299 GB
  17. MiniMax M3

    MiniMax's 428B multimodal mixture of experts with 23B active parameters: sparse attention for a 1M context, plus image and video input. The community license asks for a credit line and a notice, and written authorization above US$20 million a year.

    • MiniMax
    • 428B (23B active)
    • 1M tokens
    • MiniMax Community (Conditional)
    • from 295 GB
  18. K2 Horizon

    IFM's fully open family: a 375B mixture of experts with 23B active parameters, plus 36B-A4B, 32B, 7B, 3.7B and 0.9B models, all Apache 2.0 with a 512K context. Weights, training code, data recipes, checkpoints and logs are public, and in IFM's own table the 375B trails MiniMax-M3 on several tests.

    • Institute of Foundation Models (MBZUAI)
    • 379B (23B active)
    • 512K tokens
    • Apache 2.0 (Permissive)
    • from 0.745 GB
  19. K-EXAONE 2.0

    LG's 750B mixture of experts with 37B active parameters and a 262K context, under Apache 2.0. It is Korea's largest model, but in LG's own table it trails Qwen3.5, GLM-5.1 and DeepSeek-V4 Pro on most tests.

    • LG AI Research
    • 750B (37B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 518 GB
  20. A.X K2

    SK Telecom's 688B mixture of experts with 33B active parameters, native FP8 weights and a 256K context, under Apache 2.0. It is strongest at math and Korean, reads text only, and needs a multi-GPU node.

    • SK Telecom
    • 688B (33B active)
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 438 GB
  21. 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. In IFM's own Arabic tests the 375B scores 84.0 against 82.0 for Qwen3.5-397B, and it is trained to edit Arabic PDFs, decks and spreadsheets through file tools.

    • Institute of Foundation Models (MBZUAI)
    • 379B (23B active)
    • 512K tokens
    • Apache 2.0 (Permissive)
    • from 0.745 GB
  22. Command A+

    Cohere's open 218B enterprise model: 25B active parameters, 48 languages, images and tool use, under Apache 2.0. It runs on two H100 GPUs at 4-bit.

    • Cohere
    • 218B (25B active)
    • 128K tokens
    • Apache 2.0 (Permissive)
    • from 145 GB
  23. Nemotron 3.5

    NVIDIA's open Nemotron 3.5: a 120B Super VL, a 74B Bolt squeezed into 24 GB, and a small E4B, with a 1M context under the permissive OpenMDW license.

    • NVIDIA
    • 120B (12B active)
    • 1M tokens
    • OpenMDW-1.1 (Permissive)
    • from 5.52 GB
  24. Qwen3.8-27B

    The practical pick for a single GPU: a 27B dense multimodal model under Apache 2.0 that reads images and video and reasons step by step. By our estimate it needs about 19 GB at 4-bit.

    • Alibaba Qwen
    • 27.8B
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 19 GB
  25. Nanbeige4.3-33B

    BOSS Zhipin's 33B dense model for long agent tasks, with a 512K context, under Apache 2.0. By our estimate it needs about 23 GB at 4-bit, so it fits a 32 GB machine, but the repositories have no LICENSE file to confirm the license.

    • Nanbeige LLM Lab
    • 33.1B
    • 512K tokens
    • Apache 2.0 (Permissive)
    • from 23 GB
  26. Gemma 4

    Google's Apache 2.0 family, from phone-sized E2B models up to a 31B flagship, with reasoning, vision and tool use.

    • Google DeepMind
    • 30.7B
    • 256K tokens
    • Apache 2.0 (Permissive)
    • from 3.52 GB
  27. gpt-oss

    OpenAI's 2025 Apache 2.0 reasoning models: 117B and 21B mixtures of experts that fit one 80 GB GPU or a 16 GB machine. Older than the rest, but supported by nearly every tool.

    • OpenAI
    • 117B (5.1B active)
    • 128K tokens
    • Apache 2.0 (Permissive)
    • from 15 GB

How we chose

  • Weights you can download today, under a license we have read in full.
  • Recent releases first, plus older models that remain widely supported.
  • We do not rank by a single benchmark. Developer-reported scores are self-reported, so we describe what each model is built for and what it needs to run.
  • Every entry is checked against primary sources and re-verified when this page is updated.