Best open-source multimodal models in 2026 (images, video and audio)
Most of the new open flagships on this site read images, and many also take video or audio. This list picks the open models where that is a main feature, from phone-sized to server-sized, and shows which licenses let you ship them.
-
Gemma 4
Google's Apache 2.0 family, from phone-sized E2B to a 31B flagship. It takes text, images, video and audio, and has reasoning and tool use.
-
Qwen3.8-27B
A 27B dense Apache 2.0 model that reads images and video, reasons step by step and fits on one GPU.
-
Qwen3.6
Alibaba's Apache 2.0 Qwen3.6 pair, a 35B mixture of experts with 3B active and a 27B dense model, both with image and video input and tuned for agentic coding.
-
Qwen3.5
Alibaba's Apache 2.0 family from 0.8B to 397B, with image and video input, 201 languages and a 256K context.
-
MiniCPM-V 4.7
OpenBMB's Apache 2.0 phone-friendly models: a 35B-A3B mixture of experts and a 1B model, both with images and video.
-
MiniCPM-o 4.5
OpenBMB's 9B Apache 2.0 model that sees, listens and speaks at once. It takes video and audio live while it answers in text and English or Chinese speech, and OpenBMB lists 11 GB of GPU memory at int4.
-
Mage-VL
Microsoft's 4B Apache 2.0 vision-language model that reads video like a codec, cutting visual tokens by more than 75%, with a gate for live streams. Microsoft's tables put it ahead of Qwen3-VL-4B on most video and spatial tests.
-
LocateAnything-3B
NVIDIA's 3B model that answers with boxes and points for objects, GUI elements and text, at 12.7 boxes per second on an H100 by NVIDIA's measure. The license is research only.
-
Ternary Bonsai 2
PrismML's vision-language models with ternary weights, from a 2B model in 0.43 GB to a 27B-class model in 5.95 GB, under Apache 2.0. They take images and text.
-
SenseNova-Loop-9B-Preview
SenseTime's 9B looped vision-language preview under Apache 2.0: layers 4 to 27 run twice, with text and image input and a 256K context. It needs custom code, and its published scores are mixed against Qwen3.5-27B.
-
Muse Glimmer
Meta's 30B dense multimodal model for local agents, distilled from Muse Spark under Apache 2.0 and quantized to fit a 24 GB GPU.
-
GLM-5.3-Flash
The MIT-licensed multimodal GLM: 320B total and 18B active, images and video in, a 1M context, built for cheap long-context serving.
-
Step 3.7 Flash
StepFun's 198B vision-language mixture of experts under Apache 2.0, with 11B active parameters, a 256K context and a 1.8B vision encoder. It targets agents that read screenshots, charts and documents, and a 111.5 GB GGUF fits a 128 GB Mac or mini PC.
-
Step 5 Preview
StepFun's 600B mixture of experts with a 1M context and image input, for agents that analyze long reports, charts and screenshots. The model card also mentions video, but its specification lists only text and image, and the custom license has hardware and hosting terms.
-
MiMo-V2.6
Xiaomi's omnimodal agent models: a 1T Pro, a 310B Flash and a 9B research checkpoint, with a 1M context. The license is tagged MIT.
-
Nemotron 3 Nano Omni
NVIDIA's 31B mixture of experts with 3B active that takes video, audio, images and text with a 256K context. It's English only, answers in text and runs on one 32 GB RTX 5090 in NVFP4.
-
Qwen3-Omni
Alibaba's 30B mixture of experts with 3B active that reads text, images, audio and video and can answer in text or speech in 10 languages. Apache 2.0, but Alibaba lists about 79 GB in BF16 for a 15-second clip.
-
Muse Spark Open
Meta's 816B mixture of experts with 42B active and a 1M context, taking text, image, video and audio. It uses Meta's own Muse license, so read it before commercial use.
-
Apertus 1.5
Swiss AI's fully open 8B and 70B models under Apache 2.0, which read images and audio as well as text with a 262K context. They answer in text only, and the audio understanding is experimental.
-
LongCat-2.5 Newest
Meituan's MIT-licensed 1.6T mixture of experts with 48B active parameters, taking text, images and video with a 1M-token context. Deployment is documented only for NPU servers.
-
LFM3
Liquid AI's family has a 1.7B text model with a 256K context that fits in 1 GB, plus vision, audio and omni siblings. The LFM Open License allows commercial use only below US$10 million in yearly revenue.
-
Mirage
Cactus Compute's 51M Apache 2.0 vision model in a 22 MB file: it describes a photo, finds objects, reads text and fills forms, in 74 ms on a Galaxy S26 NPU. It is not a chat model, and its launch post gives no standard accuracy benchmarks.
How we chose
- Models that accept images, video or audio as input, with open weights you can download today.
- Current releases only, so models replaced by a newer generation, such as Qwen3-VL, are left out.
- Vision benchmark scores come from the developers, so we do not rank on them. Try your own images, because screenshots, charts and handwriting behave very differently from photos.
- Context length and license are shown because they limit what you can ship.