ollama

Key Changes

Performance Improvements

• Response speed for the /api/tags endpoint has been significantly improved in environments with large model libraries. Benchmarks show an improvement from approximately 3.1 seconds to 294ms on cold starts, and model capabilities are now reported consistently.
• Memory handling for MLX has been improved in Apple Silicon environments.

Feature Changes and Breaking Changes

• The ollama create command using MLX safetensors is no longer an experimental feature.
• When generating GGUF models, llama.cpp tools must be used for safetensor conversion and quantization.

Deprecations and Removals

• The typical_p parameter has been deprecated. It cannot be set when creating new models, and support is maintained only for existing GGUF models.

Bug Fixes and Others

• The logic for detecting token repetition runaway has been fixed, and the threshold for repeated token count has been increased to 100 to reduce false positives (e.g., during OCR tasks).
• MLX and llama.cpp libraries have been updated.