Ollama: Thinking Models Leak Control Tokens
Two independent pull requests fix the same underlying bug — user messages containing literal special-token text like "</think>" can corrupt chat templates for thinking models — while three smaller fixes target download reliability and API edge cases.
Duration: PT2M32S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-07-20T13:02:32Z
- Audio duration: PT2M32S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
It's July 20th, 2026, and this is your Ollama briefing.
The top story today is a security and correctness bug that two different contributors found and fixed independently on the same day. If a user's message happens to contain the literal text of a control token — something like "</think>" — the model's chat template can misread that plain text as an actual structural…
Second theme: pull reliability. TheJhyeFactor filed two related fixes. PR 17259 closes a gap in the download stall detector — previously, if a connection opened but never delivered a first byte, the timeout monitor never started, leaving downloads stuck indefinitely. PR 17260 adds retry logic for manifest pulls that…
Smaller fixes: deepshekhardas shipped two UX and API polish items — PR 17268 stops the sidebar from animating open on initial load, and PR 17267 maps an OpenAI "reasoning_effort" value of minimal to low instead of rejecting it outright, avoiding an unnecessary four-hundred error for a rea
sonable input.
What to remember: if you're handling chat templates or building on the OpenAI-compatible endpoint, review both special-token PRs before merging — they may overlap or conflict. And if you've seen mysterious…
Nearby episodes from Ollama
- 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
- Agent Package Gets a Major Consolidation
- Agent Hardening and a Path Traversal Fix
- Resource Leaks and GPU Placement Get a Clean-Up Pass
- Cleaning Up Memory Planning and the Launch Experience
- Fixing What Defaults Got Wrong