●Change8

v0.1.44-beta

Breaking Changes
📦 unslothView on GitHub →
⚠ 1 breaking✨ 17 features🐛 6 fixes🔧 4 symbols

Summary

This release introduces support for the new Gemma 4 12B model and significantly enhances Unsloth Studio with MCP live tools, Projects, and an experimental Canvas for interactive HTML artifacts. Installation methods have been updated, and numerous runtime fixes address CUDA compatibility across Windows, Linux, and ARM64.

⚠️ Breaking Changes

  • DO NOT USE `unsloth studio update` as packaging will not get the latest updates; use the provided curl/irm scripts for installation/updates instead.

Migration Steps

  1. Use the provided curl or irm scripts to update Unsloth or install Unsloth Studio; DO NOT use `unsloth studio update`.

✨ New Features

  • Introduction of Gemma 4 12B model support, runnable locally on 8GB RAM, supporting image, audio, and 256K context.
  • MCP (Model Composer Platform) now allows giving models live tools via one-click composer options, including built-in presets like Context7 (docs/code search), Exa (web search), and Hugging Face search.
  • MCP supports adding custom remote (OAuth/headers) or local (stdio) servers, toggled per chat.
  • New Chat UI consolidates Projects, Canvas, MCP, and Compare into a single '+\ ' menu.
  • Projects feature allows grouping related chats into a single workspace.
  • Experimental Canvas/Artifacts feature opens generated HTML in a dedicated panel inside Unsloth Studio, supporting interactive outputs and switching between preview/source code.
  • CUDA 13.3 llama.cpp binaries now work on Windows, fixing the CUDA 13.2 gibberish-output bug (default remains pinned to CUDA 13.1 for now).
  • Windows prebuilt installs no longer block on the early CUDA Toolkit check.
  • Linux llama.cpp prebuilts now match the runtime's cudart major version.
  • Prebuilt coverage extended for Blackwell (with CUDA 13.0 driver fallback) and B300 (sm_103).
  • ARM64 Linux now source-builds on GPU hosts, with a CPU prebuilt fallback.
  • ROCm detection forwards the detected AMD gfx arch to the prebuilt installer.
  • Connected models now work in Compare mode.
  • Upload limits for training datasets and recipes have been increased.
  • Window size and maximized state now persist across Studio launches.
  • Chat search now hides non-matching threads.
  • Empty content detection added for conversation datasets.

🐛 Bug Fixes

  • Fixed Apple Silicon installs that were resolving torch against x86_64.
  • Fixed UnicodeEncodeError when printing emoji on legacy Windows consoles.
  • Fixed issue where the reset-password hint did not work correctly on Windows, macOS, and Linux.
  • Fixed double-quoting of the reset-password hint for paths containing spaces.
  • Model loading now handles mid-refresh cancellation cleanly.
  • Fixed rendering for generated image frames and Python tool code blocks for cleaner presentation.

Affected Symbols

llama.cpptorchtrainerconfig