Starstruk Logo MarkStarstruk

4.46 v0.1.807-beta

medium risk

unslothai/unslothSep 8, 2026 → Sep 8, 20260 days between releases

What's New

The older release notes provided are for SeaweedFS, while the newer release is for Unsloth; they are different software products.

Breaking Changes

  • Windows code integrity blocks may prevent model loading and require reinstallation
  • AMD users relying on ROCm must switch to Vulkan as it is now the default

New Features

  • DeepSeek Harness integration added to the 'unsloth start' command
  • Studio now accepts subtitle, caption, and common attachment formats
  • Chat markdown rendering now supports on-disk sandbox images
  • Studio sliders for Min P and Repetition Penalty are now functional
  • DoRA training is now supported on Apple Silicon

Improvements

  • AMD uses Vulkan by default, delivering a 20% performance boost for prefill and decoding
  • Windows llama-server.exe is now digitally signed to reduce security scanner false positives
  • Updated PyTorch to version 2.11 bringing performance improvements and MLX enhancements
  • New Docker images published for NVIDIA GPU hosts supporting Ampere through Blackwell
  • MLX improvements include faster gated-delta models and reduced prompt memory usage

Bug Fixes

  • Over 200+ bug fixes including AppImage lag reduction, RAG upload fixes, and API issues
  • AMD Strix and iGPU gibberish issues fixed
  • MLX fixed when self healing and updating to prevent slowdowns

Upgrade Notes

Upgrade to PyTorch 2.11; AMD users should verify Vulkan compatibility; Windows users may need to address code integrity blocks.

The Two Releases

4.46

4.46

Sep 8, 2026· 1 month ago

SeaweedFS 4.46 improves S3 API (Object Lock, versioning) and volume operations while cleaning up the RDMA sidecar prototype and removing unused AI tables.

Full release notes
v0.1.807-beta

Large Perf Improvements + Fixes

Sep 8, 2026· 1 month ago

This release focuses on performance and reliability improvements for AMD and Windows users, alongside a major dependency bump to PyTorch 2.11. It introduces new Docker images, fixes over 200 bugs, and enhances Studio features including DoRA training on Apple Silicon and improved attachment handling.

Full release notes