Local vision in 2026
Vision models accept images alongside text and answer questions about them. The 2026 local landscape is good enough for OCR, chart reading, screenshot analysis, and most "describe this image" tasks. It is still behind the current frontier cloud models on the hardest visual reasoning. Deliberately unnamed: whichever three you would list today will be wrong within a year, and the gap is what matters, not the leaderboard.
| Model | Sizes | Strength |
|---|---|---|
| Llama 3.2-Vision | 11B, 90B | Strong general vision; well-supported in Ollama |
| Gemma 3 | 4B / 12B / 27B | Strong quality-per-size on modest hardware — check Gemma 4 before you settle on it |
| Qwen 2.5-VL | 3B / 7B / 32B / 72B | Structured data extraction, OCR, multilingual — strong, though no longer the only serious OCR option |
| Qwen 3.5 | 0.8B–122B local; 397B cloud-only | Vision + tools + thinking — the most capable single model here |
| LLaVA | 7B / 13B / 34B | The OG open vision model; still serviceable |
| MiniCPM-V | 8B | Compact, capable, good for edge deployment |
| Mistral Small 3.1 | various | Vision support added in 3.1 |
Picking by use case
- OCR / structured extraction: Qwen 2.5-VL, or a dedicated OCR model such as
glm-ocr. Benchmark both on your own documents; the ranking here moves faster than any list survives. - General "describe this": Gemma 3 12B/27B or Llama 3.2-Vision 11B.
- Tight memory: Gemma 3 4B or MiniCPM-V 8B.
- Vision + tools + thinking: Qwen 3.5 — already in the Ollama library, 0.8B through 122B for local tags, so pick the size your machine fits. The 397B exists only as
qwen3.5:397b-cloud.
Always verify capability
Vision support is binary at the model level — either the model has the visual encoder or it doesn't. The rows under ollama show MODEL's Capabilities header are the source of truth — a header, then one capability per row.