Ollama: MLX Comes of Age, Cancellation Gets Fixed
Ollama shifted model creation to a server-side MLX pipeline while retiring in-tree GGUF conversion, and a cluster of fixes closed out contexts and cancellation paths that could leave processes hanging or retrying forever.
Duration: PT2M49S
Episode overview
This episode is a short developer briefing from Ollama.
It explains recent repository work in plain language.
- Show: Ollama
- Published: 2026-09-15T13:02:16Z
- Audio duration: PT2M49S
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 September 15th, and today's Ollama codebase shows two clear pushes: MLX is now a first-class creation path, and a wave of fixes closed off places where the app or server could hang instead of failing fast.
Start with MLX. PR 14969, from Daniel Hiltgen, is the big one — safetensors imports now run through a server-side MLX pipeline, with upload staging, cancellation, and manifest handling built in. GGUF creation is now limited to wrapping existing GGUF files; the old safetensors-to-GGUF converter and server-side…
Second theme: cancellation and timeout discipline. Four separate PRs from DevChiniwala tighten up places that ignored context cancellation or lacked retry limits. PR 18455 and PR 18454 fix "wait for server" logic in the command line tool and the app UI so a canceled startup returns immediately instead of blocking…
Smaller but notable: PR 18443 adds support for Q2_0 GGUF tensors, PR 18462 fixes a WordPiece tokenizer bug mangling Unicode punctuation, and PR 18438 fixes case-matching logic causing intermittent "model not found" errors.
What's next: watch for the deferred MLX-C thread-local stream APIs mentioned in 18449, and expect more polish on the new MLX…
Nearby episodes from Ollama
- 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
- Tightening Up the OpenAI Compatibility Layer