What is a vision-language model, and can it run locally?
Models that read images and answer in text, and the memory they need.
Short answer
A model that reads an image and answers in text. Yes, it runs locally at the right size.
- On a 24 GB card: 6 of the 16 rated vision models with a page here run well (Q4 estimate).
- Highest rated of those: Qwen3.8 27B.
- Images use memory: Each image adds to the conversation, so several large images need more.
LMArena vision release 2026-10-19 (CC BY 4.0); memory by the site's Q4 rule on RTX 4090 24 GB. Methodology
What it is
A vision-language model is a chat model that also accepts pictures. A vision encoder turns the image into a sequence of tokens, and the language model reads those tokens together with your question. That lets it describe a photo, read a screenshot or a scanned page, answer questions about a chart, or extract fields from a receipt. It answers in text; it does not draw. Drawing is what image-generation models do, and they follow different memory rules.
Some are dedicated vision lines, such as Molmo or Qwen-VL; others are general models that read images too, such as Gemma 4. The smallest rated one with a page here, Molmo-7B-D-0924, has about 8 billion parameters and fits an 8 GB graphics card by the site's estimate.
Rated models
LMArena vision 2026-10-19: blind human votes on hosted models, a rating with its interval. Q4 need includes each machine's reserve.
| Model · parameters | LMArena vision rating | RTX 4090 · 24 GB VRAM (GB) | MacBook Air 13-inch M4 · 32 GB unified (GB) |
|---|---|---|---|
| GLM-5.3 Flash Z.ai · 320B | 1,271 ± 6 | 242 · Needs more memory | 248 · Needs more memory |
| Gemma 4 31B Google · 30.7B | 1,262 ± 6 | 26 · Runs slowly | 32 · Runs well |
| Qwen3.8 27B Alibaba · 27B | 1,242 ± 8 | 23 · Runs well | 29 · Runs well |
| Gemma 4 26B A4B Google · 25.2B | 1,242 ± 6 | 21 · Runs well | 27 · Runs well |
| MiniMax M3 MiniMax · 428B | 1,233 ± 6 | 323 · Needs more memory | 329 · Needs more memory |
| Qwen3-VL-235B-A22B-Instruct Alibaba · 235.7B | 1,215 ± 7 | 179 · Needs more memory | 185 · Needs more memory |
| GLM-4.5V Z.ai · 107.7B | 1,153 ± 12 | 83 · Needs more memory | 89 · Needs more memory |
| Llama-4-Maverick-17B-128E-Instruct Meta · 401.6B | 1,145 ± 9 | 304 · Needs more memory | 310 · Needs more memory |
| Llama-4-Scout-17B-16E-Instruct Meta · 108.6B | 1,128 ± 10 | 84 · Needs more memory | 90 · Needs more memory |
| Qwen2.5-VL-32B-Instruct Alibaba · 33.5B | 1,119 ± 15 | 28 · Runs slowly | 34 · Needs more memory |
| Molmo-72B-0924 Ai2 · 73.3B | 1,045 ± 13 | 57 · Needs more memory | 63 · Needs more memory |
| Llama-3.2-90B-Vision-Instruct Meta · 88.6B | 1,032 ± 9 | 69 · Needs more memory | 75 · Needs more memory |
Ratings: (LMArena 2026-10-19 · vision, read Oct 25, 2026)
Memory and images
The weights are planned like any text model of the same size. Each image then adds tokens to the conversation, and those tokens take room in the context cache like words do, so several large images in one chat need more memory than the plan shows. Image input also depends on the runtime: some ship the vision part as a separate file (a projector) that must be downloaded next to the model.
Side by side: Gemma 4 26B A4B vs Qwen3-VL 30B A3B for images · Molmo 7B vs MiniCPM-V 4.5 for images