Hey guys, we're super excited to launch Unsloth Studio, a new open-source web UI to train and run LLMs.
Blog + everything you need to know: https://unsloth.ai/docs/new/studio
- Run models locally on Mac, Windows, Linux
- Compare and battle models side-by-side
- Train 500+ models 2x faster with 70% less VRAM
- Supports GGUF, vision, audio, embedding models
- Self-healing Tool calling / web search + code execution
- Auto-create datasets from PDF, CSV, DOCX
- Export models to GGUF, safetensor and more formats
UnslothStudioLaunch45-4-3.mp4
What's Changed
- Update CODEOWNERS for studio and cli by @danielhanchen in #4266
- [Feature] Support Sequence Classification by @danielhanchen in #4264
- [Feature] VLMs support for GRPO by @danielhanchen in #4265
- [Fix] Respect llm_int8_skip_modules for VLM by @danielhanchen in #4249
- ROCM support by @danielhanchen in #4271
- Remove Blackwell flex attention disable workaround from studio by @danielhanchen in #4273
- ROCM support by @danielhanchen in #4272
- fix: prevent ai-assist model config RCE via untrusted Hugging Face repos by @danielhanchen in #4274
- fix(seed): disable remote code execution in seed inspect dataset loads by @danielhanchen in #4275
- Update CODEOWNERS by @danielhanchen in #4279
- fix: install data-designer plugin non-editable for Colab compatibility by @LeoBorcherding in #4268
- Arch/mixtral by @danielhanchen in #4283
- Improve documentation on how to export model from Colab by @danielhanchen in #4284
- feat: Add Mixtral model support by @danielhanchen in #4285
- Initial changes: Refactor Attention by @danielhanchen in #4286
- patch vlm trainer to resize images by @danielhanchen in #4287
- [WIP] add support for mixtral by @danielhanchen in #4288
- studio: speed up setup -- uv for installs (8x), Ninja for llama.cpp (1.7x) by @danielhanchen in #4289
- fix: remove old comments by @Shine1i in #4292
- PR: Windows Setup Improvements by @rolandtannous in #4299
- miscallenous studio by @Shine1i in #4293
- Fix: Compare Mode Deadlock, Cancel Event Poisoning & IPC Optimization by @rolandtannous in #4303
- studio: fix GGUF inference -- reasoning tokens, max_tokens, server flags, GPU allocation by @danielhanchen in #4290
- chat only with gguf for mac devices by @Manan17 in #4300
- studio: add max steps and epochs toggle switch by @Imagineer99 in #4296
- Fix/colab plugin editable install by @LeoBorcherding in #4281
- Graceful shutdown on Windows (signal handlers for Ctrl+C) by @rolandtannous in #4306
- studio: simplify auth UX to password-only login by @Imagineer99 in #4305
- studio: preserve save_steps when toggling to epochs mode by @Imagineer99 in #4308
- Fix studio frontend build producing empty Tailwind CSS by @danielhanchen in #4311
- Fix setup.sh crash on Mac with empty gitignore array by @danielhanchen in #4313
- [Feature] studio: user can upload eval dataset by @Manan17 in #4307
- fix: Ctrl+C not terminating backend on Linux by @rolandtannous in #4316
- Add download progress bar for non-GGUF models in Chat by @danielhanchen in #4314
- Apply use_reentrant removal to all TRL trainer configs by @danielhanchen in #4321
- Fix VLM GRPO matmul shape mismatch in _get_per_token_logps_and_entropies by @danielhanchen in #4301
- Improve AI Assist: Update default model, model output parsing, logging, and dataset mapping UX by @rolandtannous in #4323
- studio: per-model inference defaults, GGUF slider fix, reasoning toggle by @danielhanchen in #4325
- fix: Resolve CUDA toolkit mismatch on multi-CUDA Windows systems by @rolandtannous in #4324
- [pre-commit.ci] pre-commit autoupdate by @pre-commit-ci[bot] in #4332
- Fix/colab comment edits by @LeoBorcherding in #4317
- fix: add Qwen3.5 version gate in loader dispatch by @danielhanchen in #4335
- Fix xformers Blackwell guard: broader coverage and root cause docs by @danielhanchen in #4338
- studio: improve Colab notebook, redesign ready popup, and clean up install output by @LeoBorcherding in #4339
- Add check to disable xformers on newer GPUs by @pluesclues in #4342
- studio: training progress, CUDA lib path, dataset_num_proc fix by @danielhanchen in #4336
- studio: fix stale GGUF metadata, update helper model, auth improvements by @danielhanchen in #4346
- studio: show "Off" for repetition penalty = 1 by @danielhanchen in #4349
- studio: update Creative/Precise presets, show "Off" for disabled samplers by @danielhanchen in #4350
- studio: fix slow cancellation of GGUF generation by @danielhanchen in #4352
- Fix: Remove unused
warmupToastShownvariable (TS6133) by @rolandtannous in #4353 - Studio: SVG preview, fix streaming and model selector bugs by @danielhanchen in #4354
- fix: comment out debug print statements by @rolandtannous in #4357
- fix(llm_assist): disable thinking mode for helper model JSON output by @rolandtannous in #4358
- studio: improve onboarding UX, tooltips, and training defaults by @danielhanchen in #4355
New Contributors
- @LeoBorcherding made their first contribution in #4268
- @Shine1i made their first contribution in #4292
- @Manan17 made their first contribution in #4300
- @Imagineer99 made their first contribution in #4296
Full Changelog: https://github.com/unslothai/unsloth/commits/March-2026