Change8

v0.1.464-beta

📦 unslothView on GitHub →
19 features🐛 16 fixes🔧 8 symbols

Summary

This release introduces major support for DiffusionGemma and faster Gemma 4 via MTP, alongside significant experimental features like a Model Hub and Chat with Files (RAG) in Unsloth Studio. It also updates the underlying llama.cpp backend for improved hardware compatibility and stability.

Migration Steps

  1. To update Unsloth or install a new Unsloth Studio on macOS, Linux, or WSL, use: `curl -fsSL https://unsloth.ai/install.sh | sh`
  2. To update Unsloth or install a new Unsloth Studio on Windows, use: `irm https://unsloth.ai/install.ps1 | iex`
  3. DO NOT USE `unsloth studio update` as packaging will not get the latest updates.

✨ New Features

  • Support for DiffusionGemma models.
  • Support for Gemma 4 MTP, enabling ~2x faster performance for Gemma 4.
  • Audio chat support for Gemma 4 (wav, mp3, m4a, flac, webm) in Unsloth Studio.
  • Preserve Thinking feature added for Gemma 4.
  • Introduction of a new Hub page for browsing, downloading, and managing Hugging Face models and datasets.
  • Unsloth can now detect and display models/datasets already present on the local machine.
  • Added direct Run / New Chat actions for downloaded GGUF models from the Hub.
  • Introduction of "Chat with Files" feature in Studio for RAG capabilities over user documents.
  • RAG feature supports hybrid search, citations, PDF previews, per-thread documents, and a built-in `search_knowledge_base` tool.
  • Unsloth now uses fresh, up-to-date llama.cpp prebuilts across all supported platforms (CUDA, ROCm, Windows, Linux, macOS).
  • Added an in-app "Update llama.cpp" button to update the local backend without reinstalling Studio.
  • Improved Windows / WSL AMD support and added Strix Halo ROCm support.
  • Added Blackwell CUDA selection support.
  • Local tool calling is more reliable, featuring better ordering of tool cards, reduced duplicate tool loops, and support for tool use with GGUF vision models.
  • Improved OpenAI-compatible API and Anthropic-compatible API behavior for local Studio servers, including better error reporting, token usage, stop reasons, and Claude Code compatibility.
  • Added support for Cloudflare HTTPS free tunnels for end-to-end encrypted studios.
  • Tensor parallelism is now enabled for GGUFs, providing up to +30% throughput.
  • Improved MLX support with better model labels, generation speed stats, and fixes for VLM training.
  • Added UI polish fixes across chat, menus, model picker, dark mode, import/export, and settings.

🐛 Bug Fixes

  • Fixed issues related to non-writable Hugging Face caches during training/dataset operations.
  • Fixed custom dataset mapping issues during training.
  • Fixed kwarg spacing in training files to satisfy pre-commit.
  • Fixed Studio issue where llama.cpp update banner offered a downgrade or stuck on mixed releases.
  • Fixed false RoPE 'default' warning and rope drift gate on transformers 5.
  • Fixed Studio issue where run.py was not loaded correctly for editable installs.
  • Fixed Studio issue where `llama_extra_args` were not inherited and `--no-mmproj` was not honored.
  • Fixed Studio issue where server-loaded models were not adopted before chat auto-load.
  • Fixed stale sidebar regression test to match the gap-px markup.
  • Fixed FastModel config passthrough for sequence classification.
  • Fixed subprocess output decoding as UTF-8 in save.py on Windows.
  • Fixed EmptyLogits gathering in nested payloads and Accelerate recursively_apply.
  • Fixed Studio issue guarding Apple GPU power against negative counter-reset readings.
  • Fixed step count mismatch when sequence packing is enabled.
  • Fixed uv-bytecode-timeout issue.
  • Fixed Studio issue dropping the on-disk freshness cache after a llama.cpp update.

Affected Symbols