●Change8

v0.1.50-beta

Breaking Changes
📦 unslothView on GitHub →
⚠ 1 breaking✨ 22 features🐛 26 fixes🔧 7 symbols

Summary

This release introduces significant enhancements for AMD GPU users, enabling local LLM training and inference with improved performance and reduced VRAM usage. It also brings advanced features for running larger models, more reliable chat interactions, and enhanced web search capabilities.

⚠️ Breaking Changes

  • The `nest_asyncio` library is no longer applied on plain CLI starts. This may break asyncio behavior on Python 3.14+ if your application relies on `nest_asyncio` for certain asyncio operations in a CLI context. If you encounter issues, consider manually applying `nest_asyncio` where needed or ensuring your asyncio code is compatible with standard Python behavior.

Migration Steps

  1. If you encounter issues with asyncio in CLI applications on Python 3.14+, consider manually applying `nest_asyncio` where needed or ensuring your asyncio code is compatible with standard Python behavior, as it's no longer automatically applied.
  2. To enable the new MCP control plane, set the environment variables `UNSLOTH_STUDIO_ENABLE_MCP=1` and `UNSLOTH_STUDIO_MCP_TOKEN`.

✨ New Features

  • Local LLM training and inference on AMD GPUs across Windows, WSL, and Linux.
  • Support for training, RL, chat, and deployment of models locally on AMD GPUs.
  • Improved AMD GPU detection and installation.
  • Enhanced ROCm compatibility for AMD MI300X and MI300X GPUs.
  • Remote access to Unsloth via `unsloth studio --secure` using free HTTPS via Cloudflare.
  • Automatic GPU placement and manual selection of GPUs and model layers for running larger models.
  • Ability to move MoE expert layers into system memory to accommodate larger models.
  • Splitting models across multiple GPUs or using Tensor Parallelism.
  • Saving hardware settings separately for each model and quantization.
  • Resuming long chats without rebuilding the full conversation after an idle model frees VRAM.
  • Automatic retries for stalled Hugging Face XET downloads over standard HTTP.
  • Reusing existing GGUF files instead of re-downloading them when a model loads.
  • Web search can now read PDF papers, manuals, and other PDF results.
  • More reliable parallel tool calls, reasoning output, and tool retries.
  • An opt-in MCP (Model Control Plane) endpoint for compatible AI clients to inspect models, training history, start/stop training, load checkpoints, validate recipes, and export GGUFs. Enable with `UNSLOTH_STUDIO_ENABLE_MCP=1` and set bearer token with `UNSLOTH_STUDIO_MCP_TOKEN`.
  • Dynamic NVFP4 models that keep accuracy-sensitive layers in FP8 or BF16 while running the rest in W4A4, enabling faster inference and longer context on NVIDIA Blackwell GPUs.
  • Expanded personalization options for Studio, including color palettes, custom colors, font import, font size adjustment, contrast, and reduced motion.
  • Voice settings tab in Studio for dictation, custom dictionary, and read-aloud.
  • New display languages for Unsloth Studio: French, German, Spanish, Hindi, Arabic, Russian, and Korean.
  • Four-level tool-call permission selector for agents: Ask, Approve for me, Off, and Full access.
  • Agent workspace isolation and safer installer checks.
  • Allow API key for Ollama connections.

🐛 Bug Fixes

  • Text-only training with multimodal models no longer truncates long examples before packing.
  • Fine-tuned Qwen3.5 and Qwen3.6 MTP models now export correctly to GGUF.
  • Fixed a Windows permission error that could stop GGUF exports during the final write step.
  • Fixed `smart_chunk_text` single-chunk path leaking internal tensor type when `eos_token_id` is None.
  • Fixed Settings layout overflow.
  • Fixed nested dropdown menus staying on screen.
  • Fixed DeepSeek reasoning test shim.
  • Harden desktop release token permissions.
  • Studio: clickable sidebar settings cog, long name truncation, Canvas menu opt-in.
  • CI: opt tool-calling smoke tests out of the chat tool approval gate.
  • Fixed Inkling reasoning-effort coercion for duck-typed engine stand-ins.
  • Fixed config cards clipping content at narrow window widths.
  • Studio: use one shared Hugging Face token across Settings and training.
  • Studio: don't drop parallel tool calls after an internal no-op.
  • Studio: extract text from PDF web results.
  • Don't kill live llama-servers when a new Studio instance starts.
  • Fixed `unsloth_push_to_hub_gguf(save_method="lora")` raising NameError.
  • Relaxed Metal context-store timeout by default for MLX.
  • Fixed Studio reasoning channel rendering.
  • Fixed Studio safetensors tool reprompt reasoning leakage.
  • Propagated `fp8_block_size` before the early return in `get_lora_parameters_bias`.
  • Fixed tokenizer model check after saving, not before.
  • Stabilized Studio regression tests.
  • Skipped `.jsonl` lines that are valid JSON but not objects.
  • Fixed equal padding in the dataset source segmented control.
  • Fixed canvas staying open after starting a New Chat.

Affected Symbols

unsloth_push_to_hub_ggufsmart_chunk_textget_lora_parameters_biasunsloth studio --secureUNSLOTH_STUDIO_ENABLE_MCPUNSLOTH_STUDIO_MCP_TOKENnest_asyncio