Ollama: Hardening the Thinking and Tool-Call Pipeline
A single contributor's stress-testing of a new thinking-budget feature surfaced four separate reliability bugs in how Ollama handles long-running model responses and tool calls, while other contributors landed fixes for data races, memory sharing, and stale state across the app and server.
Duration: PT2M34S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-08-05T13:01:26Z
- Audio duration: PT2M34S
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 5th, and today's codebase activity centers on one theme: what happens when a model runs long, loops, or gets cut off mid-answer.
The headline is PR 17566 from mann1x, which proposes a token budget for model thinking — a way to stop a model from burning its entire context re-reading itself, a failure mode reported against Gemma and Qwen since last May. But the more interesting story is what happened while stress-testing that feature against a…
The second theme is state isolation. RerankerGuo landed three related fixes: PR 17573 stops chat requests from mutating a model's own message history in place, PR 17574 stops the server's log poller from duplicating device info on incomplete scans, and PR 17572 makes the benchmark tool truncate output files instead…
Also notable: Daniel Hiltgen's PR 17445 fixed a set of real data races in the progress bar and spinner code, caught by race detector testing, including one race that was live in production pull and push progress but never triggered by tests. And PR 17567 fixes a Linux build break on older glibc systems for the MLX…
What's next: watch for the thinking-budget PR 17566 to land now that its supporting fixes are…
T…
Nearby episodes from Ollama
- Closing the Gaps Where Failures Go Silent
- Streams That Fail Silently, Fixed
- Weekly Recap - Speed, Speculation, and Scheduler Reliability
- Scheduler Reliability Overhaul
- Speculative Decoding and the Cloud Model Nudge
- API Compatibility and Model Correctness Push
- Speculative Decoding Gains and a Lint Lockdown
- Cleaning Up Concurrency and Cutting Experimental Code