Best open-source vision language models in November 2026: rankings, benchmark scores, memory and serving

On the Arena vision leaderboard of 6 November 2026, the best open-weight model is Z.ai's GLM-5.3-Flash (MIT licence, 320B total, 18B active) at 1272, then Gemma 4 31B at 1265, Kimi K2.6 at 1264 and DeepSeek-V4.1-Flash at 1263. The best closed model scores 1309, so the gap is 37 points. For a single 32 GB GPU the practical picks are Gemma 4 26B-A4B (about 15 GB at 4-bit, 1242) and Qwen3.6-35B-A3B (about 20 to 23 GB, Qwen's own MMMU 81.7); on a 128 GB machine, NVIDIA's Nemotron 3.5 Super VL at NVFP4. The separate "-VL" model lines have mostly ended: the leading labs now ship vision inside their main models.

Ranks and scores from the Arena vision leaderboard, read 7 November 2026 (page dated 6 November). Ranks include closed models. Parameter counts and licences from each model's Hugging Face card. The 4-bit column is parameters x 0.5625 bytes (4 bits plus scales), my arithmetic, weights only; the Qwen3.8-27B size is ggml-org's file and the Nemotron size is from this site's Nemotron 3.5 Super page.

Which open vision models rank highest in November 2026?

Four open models sit within 10 points of each other at ranks 36 to 41, and they are all general-purpose flagships rather than specialised vision models. Arena's vision board ranks models by pairwise human votes on image prompts, which makes it the main independent comparison across vendors; GLM-5.3-Flash has 32,368 votes and Gemma 4 31B 41,989, so their ±6 confidence intervals are tight. The top 35 places are all closed models from Anthropic, Alibaba (Qwen3.8-Max), Google, Meta, OpenAI and SpaceXAI.

Below the leaders the order is: Kimi K2.5 Thinking (1252), Xiaomi's MiMo-V2.6-Flash and Pro (1248), Qwen3.5-397B-A17B (1246), Gemma 4 26B-A4B (1242), Qwen3.8-27B (1241), MiniMax-M3 (1231) and Mistral Large 4 (1230). Nemotron 3.5 Super VL, released on 15 October 2026, enters at 1211 with 5,714 votes. The last generation of dedicated VLMs is far behind: GLM-4.6V scores 1161 and Llama 4 Maverick 1146.

Two caveats. Arena measures what people prefer in a chat window, which in my view favours fluent answers over exact OCR or counting. And it ranks the model as served by its vendor or an API host, usually at native precision (Nemotron's entry is the NVFP4 build); a 4-bit copy you quantize yourself may score a little lower.

How do they score on MMMU, document and video benchmarks?

On the benchmarks the vendors publish, Qwen3.5-397B-A17B has the most complete sheet and leads most columns, and the 35B-A3B MoE models come within a few points of it. Every number below is self-reported by the model's developer, and Qwen's table is the only one that includes other open models, so read it as a Qwen-run comparison.

Read 7 November 2026. The Qwen3.6 card's table also gives VideoMMMU 83.7 and MMBench 92.8 for Qwen3.6-35B-A3B. Google's own Gemma 4 card reports OmniDocBench as edit distance (0.131 for 31B, lower is better) rather than Qwen's score, plus MATH-Vision 85.6. Kimi K2.6's card reports MathVision 87.4 (93.2 with Python). Different vendors use different prompts, thinking modes and tool settings, so compare within a row source, not across.

Three newer models report on different benchmarks altogether. Qwen3.8-27B's card leads with agent and GUI tasks: OSWorld-Verified 84.3, AndroidWorld 81.9, OmniDocBench 1.5 at 91.1, and CharXiv 90.2 with a code interpreter. Nemotron 3.5 Super VL publishes BF16 against NVFP4 scores only (MathVision 78.16 against 77.79, CharXiv reasoning 80.82 against 79.88, OCRBench v2 English 68.12 against 68.34), which says more about its 4-bit build than its rank. GLM-5.3-Flash gives one vision number in text, 90.76 on the Real5 variant of OmniDocBench.

What happened to the separate VL models?

They were absorbed. Alibaba shipped Qwen3-VL in autumn 2025, then made Qwen3.5 (February 2026) natively multimodal; the Qwen3.5, 3.6 and 3.8 checkpoints I checked on Hugging Face are all tagged image-text-to-text. Z.ai calls GLM-5.3-Flash "the first natively multimodal model in the GLM-5 series"; its earlier GLM-4.5V and 4.6V were separate vision models. DeepSeek-V4.1-Flash trains its own DeepSeek-ViT encoder jointly with the language model "from the start of pre-training", per its model card, which lists DocVQA 95.6 (4-shot).

A few dedicated lines continue. OpenBMB published MiniCPM-V 4.7 on Hugging Face in October 2026, its first MoE (35B-A3B, SigLIP2 encoder, Apache 2.0, built on Qwen3.5/3.6), plus a 1B version; its card shows benchmark charts as images only. Shanghai AI Lab's InternVL has gone quiet: the newest InternVL checkpoints in OpenGVLab's account are InternVL3.5 Flash variants from 28 September 2025, and no InternVL3.5 model appears in Arena's visible ranks. If you have a 2025 pipeline built on InternVL3.5 or Qwen2.5-VL, a 2026 MoE flagship at the same memory is the upgrade.

How much memory does a vision model need?

Weights plus the vision encoder plus the KV cache for image tokens, and the image tokens are where people underestimate. Qwen3.6-35B-A3B's config.json sets patch_size 16 and spatial_merge_size 2, so one visual token covers 32 x 32 pixels, and temporal_patch_size 2 merges pairs of video frames. Its text model has 10 full-attention layers with 2 KV heads of dimension 256; the other 30 layers are linear attention with a fixed-size state. My arithmetic at native resolution, BF16 KV cache:

KV per token = 10 layers x 2 heads x 256 dims x 2 (K and V) x 2 bytes = 20 KiB

1024 x 1024 image:   (1024/32) x (1024/32)        = 1,024 tokens -> 20 MiB of KV
1920 x 1080 frame:   60 x 34                       = 2,040 tokens -> 40 MiB
100-page scan at 1024 x 1024 per page             = 102,400 tokens -> 2.0 GiB
60 s of 640 x 360 video at 2 fps:
   120 frames / 2 (temporal merge) x (20 x 12)    = 14,400 tokens -> 281 MiB

On Qwen3.6, KV memory is small even for long video because only a quarter of its layers keep a cache. The cost is prefill compute: 102,400 tokens of scanned pages have to pass through the model before the first word appears. Real processors also resize images to a pixel budget, so check the min_pixels and max_pixels settings in your serving engine. Gemma 4 instead takes a fixed per-image budget of 70, 140, 280, 560 or 1,120 tokens, with the higher budgets recommended for OCR and small text, per its model card; video is read as frames, up to 60 seconds at one frame per second.

My recommendations from the sizes above. The 4-bit Qwen3.6 GGUF is 22.9 GB per the Qwen3.6-35B-A3B page. GLM-5.3-Flash also runs on one Spark at about 107 GB in a 2-bit build, per the GLM-5.3-Flash page, at a quality cost. Which 32 GB models fit for text is covered in the planned best LLMs for an RTX 5090.

Which engines serve open vision models?

vLLM and SGLang for anything multi-user, llama.cpp (and Ollama or LM Studio on top of it) for single-user and Mac setups. The model cards give the minimum versions: Qwen3.6 recommends vllm>=0.19.0 and sglang>=0.5.10; Nemotron 3.5 Super VL lists the vllm/vllm-openai:v0.32.0 image, TensorRT-LLM 1.3.0rc28 with a Mamba cache patch, an SGLang image and a ggml-org GGUF collection, with recipes for a single DGX Spark.

In llama.cpp the vision encoder is a separate projector file. The multimodal docs describe loading it with --mmproj in llama-server (OpenAI-compatible chat API) or llama-mtmd-cli, and list Gemma 4 among the pre-quantized models. For Qwen3.8-27B, ggml-org's repository ships the mmproj at 931 MB (BF16) or 629 MB (Q8_0) next to the 19.0 GB Q4_K_M model:

llama-server -m Qwen3.8-27B-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.8-27B-Q8_0.gguf \
  -c 32768 -ngl 99 --port 8080

The docs for vLLM on a Spark and llama.cpp on a Spark cover the build and container steps; the Spark-specific notes on image workloads are in the multimodal guide.

How fast are they?

Text generation after the image is encoded runs at the model's normal decode speed; image encoding and prefill are the extra cost, and almost nobody publishes them for local hardware. The numbers that exist:

• DGX Spark, Gemma 4 26B-A4B, text prompts, from an NVIDIA forum benchmark of 10 August 2026: 49.6 tok/s single-stream on Ollama (Q4_K_M), 54.9 tok/s on vLLM 0.26.1rc1 with a 4-bit build and MTP, and 303.6 tok/s total on vLLM at eight concurrent requests (Ollama stayed at 51.9, and the author warns the concurrent figures varied by about 30 percent between runs). Time to first token was 511 ms on Ollama and 94 ms on vLLM.
• MLPerf v6.1 Qwen3-VL-235B-A22B (published 16 September 2026): Lambda's 8x B200 did 101.56 samples/s Offline on vLLM with NVFP4 weights and FP8 KV cache; Supermicro's 8x B300 129; CoreWeave's 72-GPU GB300 NVL72 1,135. Per GPU that is 12.7, 16.1 and 15.8 by my division, against 2.25 for Red Hat's 8x H200 in the previous round (18.02 samples/s, also vLLM).

I found no published RTX 5090 timings for any current VLM. By the bandwidth arithmetic the 5090's 1,792 GB/s against the Spark's 273 GB/s should decode Gemma 4 26B-A4B several times faster once the image is in the cache; that is an inference, not a measurement. GPUwerk has not benchmarked any vision model; its Gemma 4 26B-A4B page collects the Spark reports.

Which open vision model should you use?

My picks for November 2026, by job:

• General image chat on one 24 to 32 GB GPU: Gemma 4 26B-A4B. Arena 1242, about 15 GB at 4-bit, fast because only 3.8B parameters are active, Apache 2.0.
• Documents, charts and video on one 32 GB GPU: Qwen3.6-35B-A3B. On Qwen's own table its OmniDocBench 89.9 and CharXiv 78.0 beat Gemma 4 26B-A4B's 74.4 and 69.0, and it reports Video-MME 82.5 where Gemma has no entry. For pure OCR at volume, a dedicated model may be better; see the planned local OCR model comparison.
• Screen and GUI agents: Qwen3.8-27B, with OSWorld-Verified 84.3 on its card, if you accept a dense model's slower decode.
• 128 GB box: Nemotron 3.5 Super VL at NVFP4, which NVIDIA lists for a single DGX Spark, or Qwen3.5-122B-A10B (Arena 1226) at 4-bit.
• Best open quality, budget for 256 GB or more: GLM-5.3-Flash. It is the top open model on Arena and MIT-licensed, but needs about 181 GB at 4-bit.

Whatever you pick, test it on your own images at the quantization you will run, because the published scores are full-precision and mostly self-reported. The planned open-weight model leaderboard tracks where these models stand on text benchmarks.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• Vision Arena leaderboard, Arena (formerly LMArena), 2026-11-06
• Qwen/Qwen3.6-35B-A3B model card, Hugging Face
• Qwen/Qwen3.5-397B-A17B model card, Hugging Face
• Qwen/Qwen3.8-27B model card, Hugging Face
• google/gemma-4-31B-it model card, Hugging Face
• google/gemma-4-26B-A4B-it model card, Hugging Face
• zai-org/GLM-5.3-Flash model card, Hugging Face
• moonshotai/Kimi-K2.6 model card, Hugging Face
• deepseek-ai/DeepSeek-V4.1-Flash model card, Hugging Face
• nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 model card, Hugging Face, 2026-10-15
• openbmb/MiniCPM-V-4.7-35B-A3B model card, Hugging Face
• OpenGVLab models, sorted by creation date, Hugging Face
• Qwen3.6-35B-A3B config.json, Hugging Face
• ggml-org/Qwen3.8-27B-GGUF file listing, Hugging Face
• Multimodal support in llama.cpp (docs/multimodal.md), ggml-org/llama.cpp on GitHub
• Measured inference benchmarks on a single DGX Spark across Ollama, llama.cpp and vLLM, NVIDIA Developer Forums, 2026-08-10
• MLPerf Inference v6.1: pioneering agent, VLM benchmarks, Lambda, 2026-10-22
• MLPerf Inference v6.1: Supermicro NVIDIA HGX B300 8-GPU systems, Supermicro
• MLPerf Inference v6.1 results: CoreWeave leads cloud providers, CoreWeave, 2026-10-02
• Red Hat AI tops MLPerf Inference v6.0 with vLLM on Qwen3-VL, Whisper, and GPT-OSS-120B, Red Hat, 2026-04-01