Ollama: Thinking Model Output Is Getting a Real Fix
Three separate pull requests from developer max-cyq target the same underlying problem — reasoning and thinking tokens leaking into user-facing output — across DeepSeek and Gemma models, alongside a new opt-in renderer for Qwen3.6.
Duration: PT2M33S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-07-11T13:00:37Z
- 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 11th, 2026, and this is Ollama.
Today's clearest signal: thinking-model output handling is getting a serious cleanup, with one contributor driving three related fixes in a single day.
Developer max-cyq opened PR 17128, 17127, and 17125, all addressing the same class of bug — reasoning tokens and think tags bleeding into responses users didn't ask for. PR 17128 fixes DeepSeek R1 so that setting think to false actually suppresses hidden reasoning, instead of leaking an end-of-thinking tag and raw…
Taken together, these three changes suggest the thinking-output pipeline — renderer, parser, and routing — has had inconsistent rules about when reasoning should be visible, and it's now being tightened model by model. For anyone building on top of chat completions, this matters directly: expect cleaner separation…
Related, PR 17121 adds a new opt-in renderer variant for Qwen3.6 that preserves thinking blocks by request, following the existing pattern used for Qwen3 VL — a sign the project is standardizing how thinking behavior gets toggled across model families rather than special-casing each one.
Elsewhere, smaller maintenance items: PR 17126 bumps the minimum CMake…
Nearby episodes from Ollama
- Stability Sweep Across Cloud, GPU, and Agent Tools
- Qwen3.5 Gets Untangled, and Small Bugs With Big Blast Radius
- Model Behavior Bugs and Blob Security Hardening
- Launch Gets Smarter, Model Support Gets Wider
- Cleaning Up the Scheduler's Edge Cases
- Weekly Recap - Rebuilding the Model Pipeline and Tightening the Guardrails
- Truth in Reporting
- MLX Create Pipeline Rewrite Lands