Best Open-Source Vision-Language Models (November 2026)
Best Open-Source Vision-Language Models (November 2026)
What changed as of November 2026
This page was written on July 2, 2026 and re-verified against the model repositories on November 10, 2026.
- Corrected sizes. Qwen3-VL ships as 2B, 4B, 8B, 32B, 30B-A3B and 235B-A22B (not 7B/72B — those are Qwen2.5-VL). Llama 4 is Scout (109B total, 17B active) and Maverick (400B total, 17B active); there is no 90B Llama 4 (90B was Llama 3.2 Vision). DeepSeek’s dedicated vision line is DeepSeek VL2 (Tiny, Small and base MoE, December 2024) under the DeepSeek model license — not MIT, and there is no 33B.
- New since July: Molmo 2 (Ai2, 4B/8B, Apache 2.0) succeeds the 2024 Molmo; Qwen3.8-27B (August 2026, Apache 2.0) is natively vision-language; NVIDIA Nemotron 3.5 Super VL (October 2026, 120B MoE / 12B active) and Muse Spark Open (Meta, October 2026, 816B MoE) are large open multimodal models; Kimi K3.1 Pro (Moonshot, open weights November 2026, 2.8T MoE) is natively multimodal. DeepSeek’s newest vision support comes through DeepSeek-V4.1-Flash (MIT), not a separate VL model.
- Ollama: the library carries qwen3-vl, qwen3.8, gemma4 and llama4, but not Molmo or DeepSeek VL2 — this page previously said all five run in Ollama.
- Closed comparison updated: the frontier closed models are now GPT-6 Astra, Claude Opus 5.5 / Fable 5.2 and Gemini 4 Argon.
The short answer
There is no single best open VLM — pick by job. Qwen3-VL for OCR and documents, Qwen3.8-27B for the best single-GPU multimodal all-rounder, Molmo 2 for fully open reasoning and pointing, Nemotron 3.5 Super VL or Kimi K3.1 Pro if you have server GPUs and want the largest open multimodal model, and MolmoMotion for 3D motion. Original July 2026 framing: Molmo for reasoning, Qwen3-VL for OCR, Llama 4 for scale, DeepSeek VL for cost, and MolmoMotion for 3D motion.
Last verified: November 10, 2026.
At a glance
| Model | Vendor | License | Best for | Sizes |
|---|---|---|---|---|
| Molmo / Molmo 2 | Ai2 | Apache 2.0 | Fully open reasoning, pointing, video tracking | Molmo (2024): 1B, 7B, 72B · Molmo 2 (Dec 2025): 4B, 8B, O-7B |
| MolmoMotion | Ai2 | Apache 2.0 | 3D motion forecasting | 4B |
| Qwen3-VL | Alibaba | Apache 2.0 | OCR, documents, multilingual | 2B, 4B, 8B, 32B, 30B-A3B, 235B-A22B |
| Qwen3.8-27B 🆕 | Alibaba | Apache 2.0 | Single-GPU multimodal all-rounder (images + video) | 27B dense |
| Llama 4 | Meta | Llama 4 community license | Large general VLM | Scout 109B (17B active), Maverick 400B (17B active) |
| Nemotron 3.5 Super VL 🆕 | NVIDIA | NVIDIA open model license | Documents, video, long context on servers | 120B MoE (12B active) |
| Kimi K3.1 Pro 🆕 | Moonshot | Modified MIT | Largest open multimodal agent model | 2.8T MoE (104B active) |
| DeepSeek VL2 | DeepSeek | DeepSeek model license | Cheap MoE inference | Tiny, Small, base (MoE) |
Molmo and Molmo 2 — the fully open reasoning VLMs
Ai2’s Molmo family is the closest thing to a fully open, research-transparent VLM. Weights, code, training data recipe, and evaluation harness are all Apache-licensed. Molmo 2 (December 2025: 4B, 8B and the fully open O-7B, plus video-pointing variants) is the current generation.
- Strengths: Reasoning, chart understanding, pointing (predicts pixel coordinates for referenced objects); Molmo 2 adds tracking objects across video frames
- Weaknesses: Not as sharp on OCR-heavy documents as Qwen3-VL; not in Ollama’s library
- Best for: Research, safety analysis, and anyone who wants a transparent baseline
- Sizes: Molmo 1B (MolmoE), 7B, 72B; Molmo 2 4B, 8B, O-7B
MolmoMotion — the 3D motion specialist
Released June 17, 2026. Takes a few seconds of video and predicts where objects will move in 3D over the next N seconds.
- Strengths: Language-conditioned motion, robotics-ready structured output
- Weaknesses: Not a general chat VLM — it’s a motion prior
- Best for: Robotics, autonomous driving, sports analytics, video-generation pipelines
Qwen3-VL — the OCR and document champion
Alibaba’s Qwen3-VL is still the strongest open model family for text-heavy visual tasks.
- Strengths: OCR, table extraction, chart parsing, layout understanding, OCR in 32 languages
- Weaknesses: More Chinese-training bias than Llama or Molmo; still solid on English
- Best for: Document processing, invoice OCR, multilingual UI screenshots
- Sizes: 2B, 4B, 8B, 32B (dense); 30B-A3B and 235B-A22B (MoE); Instruct and Thinking variants
Newer option: Qwen3.8-27B (August 2026, Apache 2.0) is a dense 27B model with a built-in vision encoder that handles images and video. It is a general model rather than an OCR specialist, but it is the strongest open multimodal model that fits a single 24GB GPU at 4-bit.
Llama 4 — Meta’s large general VLM
Meta’s Llama 4 shipped in April 2025 as a natively multimodal mixture-of-experts family: Scout (109B total, 17B active, 16 experts) and Maverick (400B total, 17B active, 128 experts).
- Strengths: Strong general chat, huge context windows (Scout)
- Weaknesses: Llama community license (restrictions above 700M MAU), heavy memory requirements; newer open multimodal models (Nemotron 3.5 Super VL, Muse Spark Open, Kimi K3.1 Pro) have overtaken it
- Best for: Teams already standardized on Llama tooling
- Sizes: Scout 109B, Maverick 400B (both multimodal)
The new large open multimodal models (October–November 2026)
- NVIDIA Nemotron 3.5 Super VL (October 2026): 120B Mamba-Transformer MoE with 12B active, text + image + document + video input, up to 1M context. Hosted from about $0.15/$0.50 per million tokens. See Nemotron 3.5 Super VL explained.
- Kimi K3.1 Pro (Moonshot, weights November 2026): 2.8T MoE / 104B active, natively multimodal, modified-MIT license. Server-class hardware only. See What is Kimi K3.1 Pro?
- Muse Spark Open (Meta, October 2026): 816B MoE / 42B active, text + image + video + audio input, Muse model license. See best hosts for Muse Spark Open and Nemotron 3.5 Super VL.
DeepSeek VL2 — the cheap MoE option
DeepSeek’s dedicated vision-language line is DeepSeek VL2 (December 2024), a mixture-of-experts family whose small active-parameter count keeps inference cheap. VL2-Tiny fits a consumer GPU.
- Strengths: Low active parameters, good quality relative to cost
- Weaknesses: Behind Molmo 2 and Qwen3-VL on reasoning-heavy benchmarks; DeepSeek model license (not MIT); not updated since 2024 — DeepSeek’s newer vision support lives in DeepSeek-V4.1-Flash (MIT, 552B MoE), which is far larger
- Best for: Cost-sensitive batch workloads where a 2024 model is good enough
- Sizes: VL2-Tiny, VL2-Small, VL2 (MoE)
Head-to-head by use case
OCR and documents: Qwen3-VL 8B > Molmo 2 8B > Llama 4 Scout > DeepSeek VL2
General visual chat (single GPU): Qwen3.8-27B > Qwen3-VL 8B ≈ Molmo 2 8B
General visual chat (servers): Kimi K3.1 Pro, Muse Spark Open and Nemotron 3.5 Super VL > Llama 4 Maverick (our judgment from release dates and vendor results; no single independent multimodal leaderboard covers all four yet)
Robotics motion: MolmoMotion (only specialist option; others describe, they don’t forecast)
Multilingual UI: Qwen3-VL > everything else
Cost per token: DeepSeek VL2-Tiny ≈ Qwen3-VL 2B > Molmo 2 4B
Fully open (weights + code + data): Molmo 2 > everything else
Local deployment
Support varies by model:
- Ollama — qwen3-vl, qwen3.8, gemma4 and llama4 are in the library; Molmo and DeepSeek VL2 are not
- LM Studio — GUI-driven local chat for the GGUF-converted models
- vLLM — production serving with high throughput, including Molmo 2
- HuggingFace Transformers — reference implementation for every model on this page
For a single 24GB consumer GPU (RTX 4090, RTX 5090, or Apple Silicon with 32GB+ unified memory), the sweet spot in November 2026 is Qwen3-VL 8B for OCR, Qwen3.8-27B (4-bit) for an all-rounder, or Molmo 2 8B for transparent reasoning and pointing.
Compared to closed VLMs
Open VLMs still lag the closed frontier models (GPT-6 Astra, Claude Opus 5.5 / Fable 5.2, Gemini 4 Argon) on the hardest reasoning tasks. But the gap is narrower than on text-only benchmarks — vision-language is where open models catch up fastest.
| Task | Best open | Best closed | Gap |
|---|---|---|---|
| Document OCR | Qwen3-VL | GPT-6 Astra / Gemini 4 Argon | Small |
| General chat + vision | Kimi K3.1 Pro / Qwen3.8-27B | Claude Opus 5.5 | Moderate |
| Complex reasoning | Molmo 2 / Kimi K3.1 Pro | Gemini 4 Argon | Moderate |
| Video understanding | Nemotron 3.5 Super VL, Molmo 2 + MolmoMotion | Gemini 4 Argon | Larger |
(July 2026 version of this table compared against GPT-5.6 Sol, Claude Sonnet 5 and Gemini 3.5 Pro; Gemini 3.5 Pro never reached Google’s public pricing page.)
Decision guide
- OCR-heavy work: Qwen3-VL 8B (single GPU) or 235B-A22B (multi-GPU)
- General chat: Qwen3.8-27B (single GPU), Molmo 2 8B (transparent), or Nemotron 3.5 Super VL / Kimi K3.1 Pro (servers)
- Cheap high-volume: DeepSeek VL2-Tiny or Qwen3-VL 2B
- Robotics or motion: MolmoMotion + downstream planner
- Research or safety analysis: Molmo 2 family (full transparency)
Bottom line
There’s no single best open VLM in November 2026 — pick by use case. Qwen3-VL for OCR, Qwen3.8-27B for single-GPU all-round vision, Molmo 2 for open-first reasoning, Nemotron 3.5 Super VL or Kimi K3.1 Pro for scale, and MolmoMotion for 3D motion. Start with Qwen3-VL 8B or Molmo 2 8B on a single GPU and specialize from there.
Sources
Opened November 10, 2026:
- Qwen3-VL model collection — Qwen on Hugging Face (sizes, Apache 2.0)
- Qwen3.8-27B model card — Qwen on Hugging Face
- Molmo 2 8B model card — Ai2 on Hugging Face
- MolmoMotion 4B model card — Ai2 on Hugging Face and Ai2 announcement (June 17, 2026)
- Meta: The Llama 4 herd (Scout 109B, Maverick 400B)
- DeepSeek VL2 model card — DeepSeek on Hugging Face
- NVIDIA Nemotron 3.5 Super VL model card — Hugging Face
- Kimi K3.1 Pro model card — Moonshot AI on Hugging Face
- Ollama library — vision models
Related: What is MolmoMotion? Ai2’s open 3D motion model · vLLM vs Ollama vs LM Studio · Best cheap vision AI model 2026