Ollama: Weekly Recap - Memory Discipline and Tool-Calling Cleanup
This week Ollama tightened memory management in the MLX runner, unified how model metadata and capabilities are read, and fixed a cluster of tool-call parsing bugs across Gemma 4 and Qwen3-Coder. Fifty pull request activity items and twenty-six additional commits landed, with a clear focus on reliability over new features.
Duration: PT3M17S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-09-14T09:06:38Z
- Audio duration: PT3M17S
Transcript excerpt
This excerpt keeps the crawler page concise. Listen to the episode or use the RSS feed for the full update.
This is the Ollama Weekly Recap for September 7th through September 14th. Fifty pull request activity items and twenty-six additional commits this week.
The strongest pattern is memory discipline. Jesse Gross's rework of the MLX bindings, in PR 18327, replaced a pin-and-sweep memory model with scoped array lifetimes — every array now belongs to a scope that frees it automatically when the function returns. On real hardware, this cut peak memory on a long…
The second theme is unifying how Ollama reads models. PR 17858, from Daniel Hiltgen, extracts GGUF metadata once per model blob into a cached file instead of maintaining two separate caches that could disagree. This also fixes inconsistent capability detection between models. In the same vein, PR 18374 raises the…
Third, tool-calling correctness saw a wave of fixes. Multiple contributors converged on Gemma 4 parser bugs this week: PR 18366 and PR 18398 and PR 18400 all address cases where tool arguments or object keys broke JSON parsing — string placeholder collisions and unquoted keys with spaces. PR 18422 fixes large…
Rounding out the week, the desktop app saw ChatGPT-focused polish — a first-use connection intro in PR 18321,…
Next…
Nearby episodes from Ollama
- 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
- 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