OPEN-WEIGHT MODELS · UPDATED NOVEMBER 2026
Best local vision LLMsPicks by GPU size, with scores, VRAM and license
A vision model reads scanned documents, charts, screenshots and photos as well as text. These are our picks for each size of hardware, with the vision scores each developer publishes, the GPU memory each model needs and its license.
Our picks by hardware size
ONE GPU · 32 TO 94 GB
Qwen3.8-27B
Qwen · About 24 GB · Apache 2.0
83.7% on CharXiv chart reasoning and 91.1% on OmniDocBench 1.5 in about 24 GB, with image and video input. Gemma 4 31B is the alternative from Google, at 76.9% on MMMU-Pro with image input only.
Qwen3.8-27B hardware requirements →TWO GPUS · ABOUT 190 GB
MiMo-V2.6-Flash
Xiaomi · About 188 GB · MIT
The smallest model listed that takes audio as well as images and video, so it suits meeting recordings and site footage. It scores 80.8% on OSWorld-Verified, which needs screen reading, but its card publishes no other vision scores, so test it on your own files first.
ONE 8-GPU SERVER · ABOUT 626 GB
Kimi K2.6
Moonshot AI · About 626 GB · Modified MIT
79.4% on MMMU-Pro and 80.4% on CharXiv. Pick it when one large model has to serve agents and vision together; on charts alone, Qwen3.8-27B scores higher in far less memory.
MORE THAN ONE SERVER · ABOUT 1.64 TB
Kimi K3.1 Pro
Moonshot AI · About 1.64 TB · Kimi K3.1 Pro License
84.1% on MMMU-Pro and 89.0% on CharXiv, the highest open-weight scores listed here. Check its license before commercial use.
Choosing a model for your team?
Tell us your users, your work and your data. We will recommend a model and the hardware to run it inside your own building with Vanik Edge OS. Free, no obligation.
Scores and hardware, model by model
Scores are the ones each developer reports on its own model card, without extra tools. Developers use different test settings, so compare models within a column loosely. A dash means the developer has not published that figure.
| Model | MMMU-Pro | CharXiv (RQ) | Documents and OCR | Video-MME | Memory, 1 user | Smallest setup for 16 users | License |
|---|---|---|---|---|---|---|---|
| Qwen3.6-27B Qwen model card | 75.8% | 78.4% | OCRBench 89.4% | 87.7% | About 20 GB | 1 NVIDIA L40S | Apache 2.0 |
| Gemma 4 31B Google model card | 76.9% | — | OmniDocBench 0.131 edit distance | — | About 20 GB | 1 NVIDIA L40S | Apache 2.0 |
| Qwen3.8-27B Qwen model card | — | 83.7% | OmniDocBench 91.1% | — | About 24 GB | 1 NVIDIA L40S | Apache 2.0 |
| Qwen3.6-35B-A3B Qwen model card | 75.3% | 78.0% | OCRBench 90.0% | 86.6% | About 24 GB | 1 NVIDIA L40S | Apache 2.0 |
| Nemotron 3.5 Super VL NVIDIA model card | — | 79.9% | OCRBench v2 68.3% (English) | — | About 80 GB | 1 NVIDIA RTX PRO 6000 Blackwell | OpenMDW 1.1 |
| MiMo-V2.6-Flash Xiaomi model card | — | — | — | — | About 188 GB | 2 × NVIDIA DGX Spark | MIT |
| GLM-5.3-Flash Z.ai model card | — | — | — | — | About 192 GB | 2 × NVIDIA DGX Spark | MIT |
| Kimi K2.6 Moonshot AI model card | 79.4% | 80.4% | — | — | About 626 GB | 8 × NVIDIA H100 NVL | Modified MIT |
| Kimi K3 Moonshot AI model card | 81.6% | 84.8% | OmniDocBench 91.1% | 90.0% | About 1.62 TB | 8 × AMD Instinct MI325X | Kimi K3 License |
| Kimi K3.1 Pro Moonshot AI model card | 84.1% | 89.0% | — | — | About 1.64 TB | 8 × AMD Instinct MI325X | Kimi K3.1 Pro License |
Most of these models need several GPUs working as one. Buying the GPUs is the easy part. Vanik Edge OS splits the model across them, keeps it running if a server fails and manages it all from one console. See how Edge OS clusters GPU servers
What the benchmarks measure
MMMU-Pro. College-level questions across many subjects that can only be answered by reading an image, chart or diagram.
CharXiv (RQ). Reasoning questions about charts taken from research papers.
Documents and OCR. OCRBench measures reading text in images. OmniDocBench measures turning PDF pages into text, tables and formulas; Google reports it as an edit distance, where lower is better, so it cannot be compared with Kimi K3’s percentage.
Video-MME. Questions about short and long videos, with subtitles.
FROM SIZING TO RUNNING
Sizing the hardware is the easy part.
Getting several GPUs to work as one system for a whole company is where most projects slow down. That is the part Vanik Edge OS does for you.
STEP 01 · THIS TOOL
Size it
Work out the GPU memory your models need and which GPUs or servers can hold them.
STEP 02 · VANIK EDGE OS
Cluster it
Edge OS joins your GPU servers into one cluster, splits a large model across them and keeps it running if a server fails. NVIDIA and AMD servers work side by side.
STEP 03 · VANIK EDGE OS
Run it
Your team gets AI apps with single sign-on and an audit trail, managed from one console, with your data staying inside your building.
Best local vision LLMs: questions
What is the best local vision LLM?
On a single GPU, Qwen3.8-27B, with Gemma 4 31B an alternative. On one 8-GPU server, Kimi K2.6. With more than one server, Kimi K3.1 Pro, which has the highest published vision scores of any open-weight model listed here.
Which local LLM is best for OCR and scanned documents?
On one GPU, Qwen3.8-27B scores 91.1% on OmniDocBench 1.5, level with Kimi K3’s published figure. Qwen3.6-35B-A3B scores 90.0% on OCRBench. See the document AI page for the full comparison.
Which open-weight models accept images?
Of the models listed here: Qwen3.8-27B, Qwen3.6-27B, Qwen3.6-35B-A3B, Gemma 4 31B, Nemotron 3.5 Super VL, GLM-5.3-Flash, MiMo-V2.6-Flash, MiMo-V2.6-Pro, DeepSeek-V4.1-Flash-1017, Kimi K2.6, Kimi K3 and Kimi K3.1 Pro. GLM-5.3, GLM-5.2, DeepSeek-V4-Pro, MiniMax-M2.5 and gpt-oss are text only.
Can a local model understand video and audio?
Video, yes: the Qwen models, Nemotron 3.5 Super VL, GLM-5.3-Flash, the MiMo-V2.6 models, Kimi K2.6 and Kimi K3 accept it. Audio is rarer; of the models here only MiMo-V2.6-Flash and MiMo-V2.6-Pro take it.
Do images need more GPU memory?
Yes. Each image or video frame becomes hundreds to thousands of tokens in the context, so a page-heavy document or a long video fills the KV cache quickly. Plan memory for the number of pages or frames per request.
Where do these scores come from?
Each developer’s own model card, linked in the table. We did not run the benchmarks ourselves, and developers use different test settings, so small differences between models are not meaningful.
