Ollama: Tool-Call Parsing Gets a Hardening Pass
Multiple independent fixes converged on the same weak spot this cycle: how Ollama's parsers and proxies handle tool-call data across models like Gemma4, FunctionGemma, Qwen, and GLM. A second thread tightened resource handling in the app proxy, scheduler, and GPU accounting to prevent stalls and memory growth.
Duration: PT2M33S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-09-10T13:01:24Z
- 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 day, and welcome to the Ollama briefing for September 10, 2026.
The clearest signal today: tool-calling reliability got serious attention, with three separate teams independently finding and fixing the same class of bug.
Gemma4's tool-call parser had a placeholder collision bug — when a tool call included 45 or more string arguments followed by an array, the model's valid response would silently vanish, logged only as a warning. PR 18366 and PR 18355 both fixed this, converging on prefixed decimal placeholders that can't collide…
The Responses API got similar care: PR 18343 stops assistant text from being dropped when it appears alongside tool calls, and PR 18348 from Parth Sareen lets compaction handle standalone function outputs — fixing a real failure case where Codex handoffs broke GLM sessions.
A second theme is resource and process hygiene. PR 18345 fixes a race where the scheduler could start loading a new model before a killed runner actually released memory. PR 18353 addresses unbounded memory growth in long conversations by finally allowing eviction of a conversation's own prefix-cache snapshots. And…
Smaller but notable: PR 18356 patches two WebP decoder…
Nearby episodes from Ollama
- Codex Integration Hardening and Streaming Reliability
- Tightening Up the OpenAI Compatibility Layer
- Closing the Gaps Between What You Configure and What Actually Runs
- Weekly Recap - MLX Hardens as a First-Class Engine
- Context Handling Gets Serious
- Codex Integration Push and Schema Hardening
- Structured Output Meets Speculative Decoding
- Hardening the MLX Engine and Fixing What Feeds the Scheduler