Ollama: MLX Sampler Overhaul and Codex Integration

Transcript

Good morning, this is your Ollama development briefing for May 14th, 2026.

Parth Sareen merged the Codex app integration pull request, adding comprehensive support for the Codex App as a launch integration. The changes span 20 files with over 3,300 lines added, implementing install, open, and configuration handling. The integration includes new UI components, launch icons, and reorders the launch interface to prioritize Claude Code, Codex App, Hermes, and OpenClaw.

Patrick Devine merged a significant rework of the MLX sampler system. The new implementation replaces the transform chain with an explicit distribution pipeline that applies penalties, top-k filtering, temperature and softmax, top-p, min-p normalization, and categorical sampling in sequence. This optimization keeps sparse token IDs and probabilities on GPU rather than full vocabulary scores, and fixes the seed parameter to ensure reproducible temperature sampling and MTP operations.

Daniel Hiltgen merged a smaller but important addition of memory trace logging to the MLX runner pipeline. This diagnostic enhancement is specifically targeting issue 16030 to help identify the root cause of memory-related problems.

The MLX-focused improvements represent substantial performance and reliability enhancements to Ollama's machine learning execution pipeline, while the Codex integration expands the platform's third-party application ecosystem.

What's next: Monitor the effectiveness of the new MLX sampler performance improvements and the memory trace logging for issue resolution. Track adoption and feedback on the new Codex App integration in the launch interface.

That's your Ollama development update. We'll be back tomorrow with the latest changes.