Ollama: Streaming Reliability and Parser Cleanup
Today's activity centers on fixing how Ollama handles broken or incomplete streams and malformed model output, from cloud proxy truncation to tool-call parsing, alongside a major rework of structured outputs on thinking models.
Duration: PT2M34S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-09-16T13:03:08Z
- 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 day, and welcome to Ollama, the daily developer briefing for September 16th, 2026.
The clearest pattern today: several fixes tighten up what happens when a stream or a model output doesn't go as planned, instead of failing silently.
Start with the cloud proxy. PRs 18469 and 18475, both from lorenzozanee, address the same root problem: when a cloud upstream truncates mid-stream, the server was logging the error but still telling the client everything succeeded. Now, a failed copy either aborts the handler or closes the downstream connection…
Related theme: tool call parsing. PRs 18467 and 18476 fix a case where a model emits something that looks like a tool call but doesn't match a known tool name — previously, that output just vanished, leaving an empty chat message. Now the buffered text is returned as content instead of dropped, closing out issue 17274.
The biggest structural change is jessegross's work in PRs 18478 and 18479, which collapses structured output generation on thinking models from two passes into one. Today, a formatted response on a thinking model runs an unconstrained generation, cancels it, then restarts under a grammar — costing a second prefill,…
A few other…
Nearby episodes from Ollama
- MLX Comes of Age, Cancellation Gets Fixed
- Consistency Fixes for Tool Schemas and Image Input
- Weekly Recap - Memory Discipline and Tool-Calling Cleanup
- Precision Fixes and Cleanup
- Silent Failures Get Loud
- Taming Memory in the MLX Runner
- Tool-Call Parsing Gets a Hardening Pass
- Codex Integration Hardening and Streaming Reliability