Skip to content

Vision

Ask about any image, locally.

Point a vision model at a screenshot or photo and ask. Qwen-VL on your Mac: private, on-device, no cloud.

What you can do

  • Ask about screenshots & photos — describe, extract, reason
  • Multi-image messages — compare or combine several
  • EXIF-aware & pixel-budgeted — correct orientation, safe memory
  • Feature caching — fast follow-ups in the same chat

Models

Open vision-language models, MLX-accelerated.

  • Qwen-VL and other MLX-VLM models
  • OpenAI image_url content blocks
  • Native media inputs
  • Text-only models safely refuse images

Powered by MLX-VLM. See the vision docs.

How it works

Just add an image to a chat request.

POST localhost:1337/v1/chat/completions
{ "model": "qwen2.5-vl", "messages": [
  { "role": "user", "content": [
    { "type": "text", "text": "What's in this image?" },
    { "type": "image_url", "image_url": { "url": "data:image/png;base64,..." } }
  ] }
] }

FAQ

Can I run vision models locally on a Mac?

Yes. Bedrova runs Qwen-VL and other vision models on Apple Silicon. Attach an image to a chat and ask about it, entirely on-device.

Can I send multiple images?

Yes. Multi-image messages work, with EXIF orientation handled and a pixel budget so a large photo does not blow up memory.

How is this different from perception?

A vision model reasons about an image in language. Core ML perception answers narrower questions with typed results: where the text is, what objects are present, how far away things are. Bedrova does both, and they are good at different jobs.

How do I use it from code?

Send OpenAI-compatible image_url content blocks to /v1/chat/completions (or /anthropic/v1/messages) with a vision model selected.