Ollama: Precision Fixes and Cleanup
Today's activity centers on correctness fixes for edge cases in model output handling, plus a disk-space bug in the quantization workflow. Three community integration listings round out the docs updates.
Duration: PT2M21S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-09-13T13:01:00Z
- Audio duration: PT2M21S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
Good day, it's September 13th, 2026, and this is your Ollama development briefing.
The clear thread today is precision — making sure Ollama handles edge cases correctly, whether that's numbers, streaming order, or disk cleanup.
Start with PR 18422. A Qwen3-Coder tool call using a very large number, one times ten to the twentieth, was silently getting corrupted into the wrong value on Mac ARM64 systems, because the parser forced every whole number into a 64-bit integer regardless of range. Now, values outside that range stay as floating…
Related in spirit is PR 18413, which fixes an ordering issue in the OpenAI-compatible Responses path. When a model uses web search, an open reasoning block needs to fully close before a client tool call gets emitted, or the item indices in the stream get scrambled. The fix keeps stream output in the same order it…
The other significant fix is PR 18424, addressing a disk space bug in `ollama create` with quantization. When converting from safetensors, the intermediate F16 blob wasn't being cleaned up after quantization produced its final layer. One reporter had 69 orphaned blobs eating over 800 gigabytes, with a single 26…
Rounding out the day: three…
Nearby episodes from Ollama
- Silent Failures Get Loud
- Taming Memory in the MLX Runner
- Tool-Call Parsing Gets a Hardening Pass
- 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