v0.1.37-beta
📦 unslothView on GitHub →
✨ 15 features🐛 24 fixes🔧 12 symbols
Summary
This release focuses heavily on enhancing the Unsloth Studio experience with a major UI/UX overhaul, new chat features, and expanded model support, alongside significant stability and hardware compatibility fixes, particularly for ROCm.
Migration Steps
- Pin bitsandbytes to continuous-release_main on ROCm for 4-bit decode fix.
- Pin kernels==0.12.1 to fix training import failure.
- Pin peft to 0.18.1 to fix export subprocess issues.
- Bump Studio installer minimum to 2026.4.5.
- Configure PyTorch mirror using UNSLOTH_PYTORCH_MIRROR environment variable if needed.
✨ New Features
- Revamped Unsloth Studio UI/UX emphasizing chat and training.
- Added a collapsible sidebar to Unsloth Studio.
- Ability to delete chats and search past conversations in Studio.
- New Preserve Thinking toggle for models supporting it (e.g., Qwen3.6).
- Expanded Settings page in Studio with options for profile picture, name, etc.
- Removed requirement to enter Hugging Face token twice.
- gpt-oss now features low, medium, and high thinking toggles.
- Now uses latest llama.cpp prebuilt, including on Linux CUDA.
- Kimi-K2.6 can now be run.
- Added experimental API support to Unsloth Studio.
- Added AMD ROCm/HIP support across installer and hardware detection.
- Studio: Expose openai and anthropic compatible external API end points.
- Studio: Add folder browser modal for Custom Folders.
- Add support for OLMo-3 model.
- Add cactus QAT scheme support.
🐛 Bug Fixes
- Only run ldconfig CUDA-linking recovery when permissions allow.
- Fix Mistral DPO/preference training crash on non-xformers platforms (e.g. Intel XPU).
- Fix raw text paragraph break normalization.
- Studio: keep chat input visible and fix compare pane clipping.
- Fix tokenizer check: check find() return value before adding offset in try_fix_tokenizer.
- Fix Gemma-4 GRPO catastrophic KL divergence with TRL 1.0.0+.
- Studio: fix route transition DOM duplication via AnimatePresence mode="wait".
- Studio: fix api-keys access + refresh.
- Studio: fix default chart view to full training history.
- [Studio] Install flash attn at setup time for linux.
- Fix: remove 300s cap on load_checkpoint (inherits 3600s default).
- Studio: honor explicit GGUF ctx and default to 4096 when weights exceed VRAM.
- Studio: make GGUF disk-space preflight cache-aware.
- Studio: anchor ctx-slider warning threshold at 4096 when weights exceed VRAM.
- Studio: hard-stop at n_ctx with a 'Context limit reached' toast.
- Fix num_items_in_batch GA for Gemma4.
- Fix bitsandbytes ROCm install by using pip instead of uv.
- Fix: prevent offline freeze by fixing stats retry and forwarding local_files_only.
- Respect classification head skip list on pre-quantized 4-bit checkpoints.
- Fix(rocm): tighten gfx regex to ignore generic ISA lines.
- Fix grad-accum accepts_loss_kwargs detection for vision wrappers.
- Fix grpo_compute_loss_slow called with wrong positional args.
- Gate trl disable_gradient_checkpointing warning on UNSLOTH_ENABLE_LOGGING.
- Studio: refresh Downloaded GGUF list and recurse into variant subdirs.