Ollama: Vision Support Lands in the MLX Runner
A multi-day effort from Jesse Gross wired full image input support into the MLX runner, spanning prompt handling, caching, and quantization, while release v0.32.7 shipped alongside smaller CLI and parsing fixes.
Duration: PT2M31S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-08-10T13:01:36Z
- Audio duration: PT2M31S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good morning. It's August 10th, and today's build reflects one focused engineering push: getting images working correctly and safely in the MLX runner.
The headline is PR seventeen thousand six hundred, which adds image input support to the MLX runner. But that single PR sits on top of a whole chain of preparatory commits. Sixty b-d-c-two-three expands image tags into placeholder tokens so a prompt can carry both text and media in order. A-f-five-b-two-seven makes…
A related theme is precision. Commit eight-seven-one-three-five-seven-zero found that quantizing vision towers badly degrades image recognition — in some cases to complete blindness — so vision and audio towers now stay at source precision instead of being compressed like language weights.
Outside of vision work, release v0.32.7 in PR seventeen thousand six hundred forty-six ships the Glimmer model, a prefill reporting fix to prevent memory grinding on media requests, and a CI reliability fix for CUDA downloads. Smaller but useful: PR seventeen thousand six hundred forty fixes dragged image paths in…
What to remember: if you're working with MLX vision models, re-import checkpoints to pick up the new precision policy, and…
Th…
Nearby episodes from Ollama
- Launch Ecosystem Expands, Parser Edge Cases Get Squashed
- Backup Collisions and a Security Fix Converge
- Fixing the Defaults That Silently Hurt You
- Model Coverage and Tool-Call Reliability
- Weekly Recap - Vision Runners and Robustness Cleanup
- Edge Case Cleanup Across the Stack
- Speculative Decoding Gets a New Draft Architecture
- Crash Fixes and the Launcher Ecosystem Grows