v0.1.462-beta
📦 unslothView on GitHub →
✨ 21 features🐛 15 fixes🔧 6 symbols
Summary
This release introduces major features including DiffusionGemma support, Gemma 4 MTP acceleration, experimental Hub and RAG capabilities, and updates the underlying llama.cpp backend for improved hardware compatibility and performance. Users must update via the provided installation scripts to receive the latest changes.
Migration Steps
- If DiffusionGemma wasn't previously working, install the latest `v0.1.462-beta`.
- To update Unsloth or install a new Unsloth Studio, use the provided installation scripts (curl for macOS/Linux/WSL or irm for Windows). DO NOT USE `unsloth studio update`.
✨ New Features
- Support for running and training DiffusionGemma models.
- Introduction of Gemma 4 MTP for ~2x faster performance on Gemma 4.
- Audio chat support for Gemma 4 (supporting wav, mp3, m4a, flac, webm).
- Preserve Thinking feature added for Gemma 4.
- New experimental Hub page for browsing, downloading, and managing Hugging Face models and datasets.
- Unsloth now detects existing models/datasets on the machine and displays them in the Hub.
- GGUF models downloaded via the Hub now feature direct 'Run / New Chat' actions.
- Experimental 'Chat with Files' feature in Studio supporting RAG, hybrid search, citations, PDF previews, and a built-in `search_knowledge_base` tool.
- Unsloth now uses fresh llama.cpp prebuilts across CUDA, ROCm, Windows, Linux, and 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 with better ordering of tool cards, fewer 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 (better errors, token usage, stop reasons, and Claude Code compatibility).
- Added Tool Call Permissions (Approve, Always Approve, Deny).
- Reduced tool call nudging issues by 50% to 90% without accuracy loss.
- MCP and Artifacts are now selectable.
- Tensor parallelism is now enabled for GGUFs, providing +30% throughput.
- Added Cloudflare HTTPS free tunnels for end-to-end encrypted studios.
- Improved MLX support with better model labels, generation speed stats, and fixes for VLM training.
- Added many UI polish fixes across chat, menus, model picker, dark mode, import/export, and settings.
🐛 Bug Fixes
- Fixed kwarg spacing in training files to satisfy pre-commit.
- Fixed false RoPE 'default' warning and rope drift gate on transformers 5.
- Fixed Studio issue where run.py was not loaded by path 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 model was 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 decoding subprocess output as UTF-8 in save.py on Windows.
- Fixed EmptyLogits gathering in nested payloads and Accelerate recursively_apply.
- Fixed Apple GPU power guard against negative counter-reset readings.
- Fixed step count mismatch when sequence packing is enabled.
- Fixed uv-bytecode-timeout issue.
- Fixed respect absolute export paths to prevent cross-drive copy failures (WinError 112).
- Fixed Studio issue where llama.cpp update banner offered a downgrade / stuck on mix releases.
- Fixed deduplicate lemonade ROCm prebuilt selection log.