v0.1.463-beta
📦 unslothView on GitHub →
✨ 22 features🐛 15 fixes🔧 7 symbols
Summary
This release introduces major features including DiffusionGemma support, Gemma 4 MTP acceleration, experimental Hub and Chat with Files capabilities, and significant backend updates to llama.cpp and hardware support. Users are strongly advised to update using the provided installation scripts rather than the 'studio update' command.
Migration Steps
- If DiffusionGemma was not working, install the latest version using the provided installation scripts (curl for macOS/Linux/WSL or irm for Windows), as using 'unsloth studio update' is discouraged.
- To get the latest updates, use the installation scripts provided (curl or irm), as 'unsloth studio update' will not fetch the latest packaging.
✨ New Features
- Support for running and training DiffusionGemma via Unsloth Studio.
- Introduction of Gemma 4 MTP, enabling ~2x faster performance for Gemma 4 models (MTP is auto-enabled in Unsloth Studio).
- Audio chat support added for Gemma 4, supporting formats: wav, mp3, m4a, flac, webm.
- Preserve Thinking feature added to Gemma 4.
- New experimental Hub page for browsing, downloading, and managing Hugging Face models and datasets.
- Unsloth can now detect locally present models and datasets in the Hub view.
- Downloaded GGUF models now feature direct 'Run / New Chat' actions in the Hub.
- Experimental 'Chat with Files' feature added to Studio, enabling Q&A over user documents.
- Chat with Files supports hybrid search, citations, PDF previews, per-thread documents, 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, Strix Halo ROCm support, and Blackwell CUDA selection.
- 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 (better errors, token usage, stop reasons, and Claude Code compatibility).
- Tool Call Permissions (Approve, Always Approve, Deny) added.
- 50% to 90% reduction in tool call nudging issues without accuracy loss.
- MCP and Artifacts are now selectable.
- Tensor parallelism enabled for GGUFs, providing +30% throughput.
- Cloudflare HTTPS free tunnels added for end-to-end encrypted studios.
- Improved MLX support with better model labels, generation speed stats, and fixes for VLM training.
- Exposed runtime context length for hub models.
- Studio now shows Apple GPU temperature and power metrics on macOS.
🐛 Bug Fixes
- Fixed an issue where DiffusionGemma might not work previously (suggests installing v0.1.462-beta).
- Fixed kwarg spacing in training files to satisfy pre-commit.
- Fixed a 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 to inherit llama_extra_args and honor --no-mmproj.
- Fixed studio issue to adopt server-loaded model 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.
- Guarded Apple GPU power against negative counter-reset readings.
- Fixed step count mismatch when sequence packing is enabled.
- Fixed uv-bytecode-timeout issue.
- Dropped the on-disk freshness cache after a llama.cpp update.
- Fixed respect absolute export paths to prevent cross-drive copy failures (WinError 112).