Ollama: MLX Performance Push and a Scheduler Race Fix
Ollama's MLX runtime got a cluster of performance and correctness fixes around mixed-precision quantization and memory residency, while a data race in the model scheduler's status handler was closed under lock discipline.
Duration: PT2M30S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-07-25T13:01:00Z
- Audio duration: PT2M30S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good day, it's July 25th, and here's what mattered in the Ollama codebase.
The clear theme today is MLX performance and correctness, with several changes converging on how models are quantized and kept in memory. Jesse Gross's PR 17336 fixed expert quantization in the qwen three five model, where mixed-precision checkpoints — say an eight-bit head paired with four-bit experts — were…
That same precision theme shows up in PR 17383, where draft model output heads now quantize at the requested type rather than an eight-bit default. Since draft quality only affects acceptance rate, not final output, this shaved real cost off every draft step — a measured twenty-six to thirty-seven percent decode…
Daniel Hiltgen's Laguna MLX work, spanning PR 17237 and a follow-up commit, adds new model support with a mixed quantization policy across dense and routed expert layers. PR 17367 builds directly on it, configuring Metal to keep model weights resident in memory instead of paging them, which was necessary to actually…
Second theme: concurrency safety. PR 17376 fixed a real data race — the model status handler was reading the scheduler's loaded-model map and runner fields without holding the…
Roundi…
Nearby episodes from Ollama
- A Server Reliability Sweep
- Tool Calling Gets a Reliability Pass
- Scheduler Race Conditions Take Center Stage
- Command Line Consolidation and Hardware Fixes
- Thinking Models Leak Control Tokens
- Weekly Recap - Hardening the Runtime, One Edge Case at a Time
- Closing the Gaps Between API Layers
- Cleaning Up the Agent Loop and Widening Model Support