v0.32.10-rc1
Breaking Changes📦 ollamaView on GitHub →
⚠ 1 breaking✨ 1 features🐛 1 fixes🔧 1 symbols
Summary
This release adjusts the default repeat penalty for models and introduces performance improvements for NVFP4 MLX models. A bug related to blob verification has also been fixed.
⚠️ Breaking Changes
- Models that previously defaulted to a `repeat_penalty` of 1.1 will now default to 1.0 (off). If older models exhibit repetition, set a per-model parameter to adjust the penalty.
✨ New Features
- Faster prefill on NVFP4 MLX models with a global scale, resulting in approximately 7-8% speedup on Qwen3.6 and Muse Glimmer.
🐛 Bug Fixes
- Blob verification is no longer skipped when an OCI manifest's config and layer share the same digest.