●Change8

v0.1.45-beta

Breaking Changes
📦 unslothView on GitHub →
⚠ 1 breaking✨ 6 features🐛 28 fixes🔧 7 symbols

Summary

This release focuses heavily on bug fixes and cross-platform support, particularly within Unsloth Studio, improving UI stability, model loading (GGUF, Gemma, Qwen), and installation procedures. Users must now use platform-specific scripts for updates instead of `unsloth studio update`.

⚠️ Breaking Changes

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

Migration Steps

  1. To update Unsloth or install a new Unsloth Studio, use the platform-specific installation scripts provided (curl for macOS/Linux/WSL, irm for Windows) instead of `unsloth studio update`.

✨ New Features

  • Studio: Enable audio input for Gemma 4 GGUFs.
  • Studio: Enable GGUF tools with vision inputs.
  • Studio: Support text-only loading of Gemma 3 27B via FastLanguageModel (skipping SiglipVisionModel).
  • Studio: Implement Hub + Download Manager.
  • Studio: Improve OpenAI- and Anthropic-compatible API spec compliance.
  • Studio: Implement bulk export and import in Settings Chat Data, and an MCP pill off switch.

🐛 Bug Fixes

  • Port KTO logps truncation guard to TRL 1.x _compute_logps refactor (and restored later).
  • Studio: Disable mlx gc for none.
  • Fix chat text cutoff at composer dock and speed up plus icon spin.
  • Studio: Refine tool call and reasoning trigger UI.
  • Fix warning when localhost resolves to ::1 but Studio is bound only to 127.0.0.1.
  • Fix Studio thread persistence on first runStart to keep the prompt during mid-stream refresh.
  • Studio: Accept system-role messages in Claude Code requests.
  • Studio: Fix load_freeze audio-type tests for Gemma 4 \<|audio|> probe.
  • Studio: Fix chat preset persistence with fast mode.
  • Studio: Fix Repo tests (CPU) by stopping ROCm test from leaking a fake utils into sys.modules.
  • Studio: Stop ROCm amd-smi tests leaking fake loggers into sys.modules.
  • Studio: Remove red border on chat error messages.
  • Studio: Keep chat in place when composer attachments resize it.
  • Studio: Stop leaking internal exceptions to API clients; harden sandbox path.
  • Fix MoE LoRA target parameter handling.
  • Fix Qwen 3.5 export issues.
  • Fix Nvfp4 load.
  • Studio: Fix recipe dataset preview.
  • Studio: Follow-up fix for GGUF developer prompts.
  • Studio: Stop the providers dialog from resetting custom provider form state.
  • Studio: Fix block arbitrary external image URLs in markdown renderer.
  • Fix clearer Studio setup error when GPU driver is too old for the installed CUDA toolkit.
  • Studio: Fix faithful conversation export and import round trips (ShareGPT system role, CSV quoted newlines).
  • Windows/WSL installer: Fix winget msstore cert failure, amd-smi DiskPart prompt, and enable AMD GPU (Strix Halo gfx1151).
  • Fix nested dropdown submenus clipped by the menu alignment nudge in Studio.
  • Fix dataset video paths validation before training.
  • Restore config use_cache in for_inference after gradient checkpointing prep.
  • Fix persist Windows ROCm BNB version.

Affected Symbols

TRL 1.x _compute_logpsGemma 4 GGUFsQwen3.5-4B-MTPGemma 3 27BSiglipVisionModelROCm amd-smi testsMoE LoRA target parameter handling