Control permissions

Unsloth, like Claude Code and Codex, provides permission controls for any model that uses tool calls. This means the model and Unsloth cannot access, modify, or edit your files or use the internet without your approval.

You can choose the level of access you want: Unsloth can run tools inside a secure sandbox environment to protect your system, or, if you prefer, you can allow it to directly access and edit your files.

Image and Video Generation

For MiniMax-H3 FP8 on an NVIDIA B200, a 960×544, 124-frame, 8-step generation dropped from 70+ seconds to 13 seconds. Generate with FLUX, Z-Image, LTX, Wan and fine-tuned LoRA adapters.

Transform, inpaint, extend, upscale, reference and edit existing images.

Code execution

Get up to 50% more accurate tool-calling with self-healing tool calls that detect, repair and retry failures automatically.

Execute Bash and Python in a secure sandbox. Unsloth offers more reliable call termination to reduce loops, better healing and deduplication to prevent XML leakage.

Access model anywhere

Serve your local or over HTTPS through Unsloth's free Cloudflare tunnel. Check a run from your phone, your laptop, or anywhere else you happen to be.

Bind the app to your network with -H 0.0.0.0, or open a free Cloudflare tunnel for HTTPS.

Use the Latest models

You can run and train nearly every model, including upcoming ones. Expect Day Zero support for models like Qwen3.8, Gemma, Meta, NVIDIA, GLM, Gemma models and more. It's all thanks to llama.cpp and Hugging Face and we're also proud to contribute back to the ecosystem.

Deep Research

Normal search utilises the web for its answers and happens while the model is still thinking.

Deep research plans first, searches for the best sources, then produces a detailed report with all citations.

Train Diffusion Models

Train LoRA adapters for SDXL, FLUX.2, Qwen-Image and Z-Image on your own images. Caption them in Studio, pick a rank, and hit train.

Train with no code

Drop in a PDF, CSV or JSON and go. LoRA, full fine-tuning, pretraining. All 2x faster, 70% less VRAM, no accuracy loss. Multi-GPU and latest models work.

Text, diffusion, audio and image models are supported!

Transcribe/generate audio

Generate, fine-tune or transcribe audio with Unsloth completely locally. Text-to-speech, speech-to-text, Whisper, Qwen3-ASR, you name it!

Use Cloud Models

Run models from OpenAI, Anthropic, Ollama, llama.cpp, vLLM, and more.

Use the same Unsloth chat interface for local and cloud models with support for tool-calling, image gen, prompt caching to reduce token usage while preserving provider-native features like OpenAI’s web search and code execution.

Frequently asked questions

Do you collect my data?

No telemetry. Unsloth detects your GPU type and device so the app can know what works. The app can run entirely offline.

Can I use models I already downloaded?

Yes, they are found automatically. If yours are not you can specify your own custom folders.

Why is inference slower sometimes?

Web search, code execution and tool-call healing all cost time. Turn them off and speed should match any other llama.cpp app. Still slow? Open a GitHub issue.

GPU only?

No. Unsloth works on a wide variety of CPU, Mac etc setups.

Does it support OpenAI-compatible APIs?

Yes. See the API guide. We also support connection to Cloud models or other APIs like Anthropic or OpenAI.

What devices does Unsloth support?

Unsloth supports all OS including Mac, Windows, Linux and WSL and supports NVIDIA, Intel, AMD and Mac GPUs/CPUs. Older hardware however may not be well supported.

A huge thank you to NVIDIA and Hugging Face for being part of our launch. Also thanks to all of our early beta testers for Unsloth Desktop, we truly appreciate your time and feedback. We’d also like to thank Jan for inspiration and llama.cpp, PyTorch, stablediffusion.cpp, and open model labs for providing the infrastructure that made Unsloth Desktop possible.