Ollama: Speculative Decoding Gains and a Lint Lockdown
Ollama's MLX engine picked up automatic speculative decoding for Qwen 3.5, delivering up to 70% faster generation on some models, while the team closed a gap in linting coverage that had been letting issues slip through code review.
Duration: PT2M33S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-07-30T13:00:49Z
- Audio duration: PT2M33S
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 July 30th, 2026, and here's what moved in Ollama's codebase.
The headline is performance. PR 17454 from Jesse Gross teaches the MLX engine to use the MTP head already present in Qwen 3.5 checkpoints as a built-in draft model for speculative decoding. It kicks in automatically, no configuration needed. On an M5 Max, the 35B mixture-of-experts model jumped from 131 to 162…
Second theme: code quality enforcement. Daniel Hiltgen shipped two paired changes, PR 17456 and PR 17457, that clean up lint issues across the entire tree and switch CI from only flagging new issues to scanning everything. The old setup meant a change in one place could introduce a lint problem somewhere else…
Smaller but worth noting: PR 17453 fixes tool parameter parsing so JSON Schema constraints like minimum, maximum, and default values survive the round trip instead of being silently dropped — a real correctness fix for anyone building tool-calling workflows. And PR 17472 extends debug request logging to also capture…
The rest of today's activity is routine README typo fixes and one community integration addition, not code-level changes.
What to remember: if you're running Qwen 3.5 through…
Nearby episodes from Ollama
- Cleaning Up Concurrency and Cutting Experimental Code
- Trust Your Cache, Fix Your Logs
- Parsing Bugs and Process Hardening
- Weekly Recap - New Model Support & Concurrency Hardening
- Closing the Gaps Between Local Config and Cloud Compatibility
- MLX Performance Push and a Scheduler Race Fix
- A Server Reliability Sweep
- Tool Calling Gets a Reliability Pass