Unsloth Desktop on Windows: the setup guide I wish I’d had

August 14, 2026 6 Min Read

Unsloth Desktop launched on August 11, 2026. It’s free, open source, and currently at version 0.1.701-beta. Treat it as beta software, because it is.

Watch Full Tutorial

First, the three-product thing

A lot of confusion online comes from this, so let’s clear it up before you download anything.

Product	What it is	Who it’s for
Unsloth Desktop	Native app (Tauri), installed from an `.exe`	Almost everyone. Start here.
Unsloth Studio	Browser web UI on localhost	Servers, remote boxes, WSL
Unsloth Core	Python package	Writing your own training scripts

The PowerShell one-liner circulating on social media installs Studio, not Desktop. If you ran it and got a browser tab instead of an app, that’s the reason.

What you need

Windows 10 or 11. NVIDIA, AMD, and Intel GPUs all work, and so does CPU-only, though it will be slow. There’s no Python, Docker, or WSL requirement for the desktop app.

Step 1: Install Unsloth Desktop

Download the Windows installer:

wget https://github.com/unslothai/unsloth/releases/download/v0.1.701-beta/Unsloth-Desktop-0_1_701_beta-Windows.exe

Or grab the current build from unsloth.ai/download, which is safer long-term since that link pins an old version once they ship an update.

Run it, then launch the app. That’s the whole install.

Optional: Unsloth Studio instead

If you want the web UI, open PowerShell and run:

powershell

irm https://unsloth.ai/install.ps1 | iex

Launch it with:

powershell

unsloth studio -p 8888

The same command updates Studio later. On an Intel GPU or anything else where you want the Vulkan backend for GGUF inference, set the variable before installing, because it picks which llama.cpp bundle gets downloaded:

powershell

$env:UNSLOTH_FORCE_VULKAN=1
irm https://unsloth.ai/install.ps1 | iex

A few other installer flags, if you need them:

powershell

# Skip PyTorch entirely (GGUF-only mode, much smaller install)
$env:UNSLOTH_NO_TORCH=1; irm https://unsloth.ai/install.ps1 | iex

# Don't auto-launch after install
$env:UNSLOTH_SKIP_AUTOSTART=1; irm https://unsloth.ai/install.ps1 | iex

# Pin a Python version
$env:UNSLOTH_PYTHON='3.12'; irm https://unsloth.ai/install.ps1 | iex

# Install somewhere other than the default
$env:UNSLOTH_STUDIO_HOME='C:\path'; irm https://unsloth.ai/install.ps1 | iex

On macOS, Linux, or WSL, substitute curl -fsSL https://unsloth.ai/install.sh | sh and check the install docs for the differences.

Step 2: Download a model

Open Model Hub. On Device shows what you already have, Discover searches Hugging Face.

If you’ve used Ollama or LM Studio before, your existing GGUF files get picked up automatically. You don’t need to re-download anything. If some are missed, point Unsloth at the folder in settings.

Models with day-zero support include Qwen3.8, Kimi K3, MiniMax-H3, Gemma 4, DeepSeek-V4, Muse Glimmer, and GLM-5.2. Pick a quantization that fits your VRAM and download it.

Step 3: Raise the context length before loading

This is the step that decides whether your first hour is good or frustrating. The default context window is small, and the app will feel broken the moment you paste in a long document.

Hover the model, open the settings gear, and raise the context length. Unsloth auto-fits GPU layers and expert offloading to whatever you choose, so the guess-and-check loop from LM Studio isn’t necessary. Tick remember for this model, then load it.

Step 4: Set your tool permissions

Unsloth runs tool calls and executes Bash and Python in a sandbox, with permission controls modeled on Claude Code and Codex. You pick the level:

Ask every time. Nothing runs without your approval.
Sandbox only. The model works in an isolated environment and can’t touch your real files.
Full access. It reads and edits your files directly.

Start restrictive. Middle ground that works well day to day: allow tool calls, require approval for high-risk actions.

Step 5: Web search and Deep Research

Both are built in, private, and don’t need an API key.

Web search runs while the model is still reasoning, so sources come in mid-thought.

Deep Research plans before it searches. It drafts a numbered research plan, lets you edit it (add steps, remove steps, tighten the scope), and only then runs and produces a report with citations.

Expect this to take a while on consumer hardware. Duration depends on your model, context length, and machine.

Step 6: Chat with your documents

Add files to a chat and Unsloth parses, chunks, and embeds them locally. PDFs, DOCX, and tables are handled, including right-to-left and Indic text. Recent builds let you change the embedding model and search Hugging Face for alternatives.

Nothing leaves your machine. Unsloth collects no telemetry and the app runs fully offline.

Step 7: Point Claude Code or Codex at your local model

Load a model in Unsloth, open your project folder in a terminal, then:

powershell

unsloth start claude

Swap the agent name for any of these:

Agent	Command
Claude Code	`unsloth start claude`
OpenAI Codex	`unsloth start codex`
Hermes Agent	`unsloth start hermes`
OpenClaw	`unsloth start openclaw`
OpenCode	`unsloth start opencode`

Claude Code, Codex, and OpenCode can also keep their existing cloud model and use your local model as a subagent for cheaper work:

powershell

unsloth start claude --as-subagent --model unsloth/model-GGUF:quant

Replace unsloth/model-GGUF:quant with the model and quantization you actually loaded.

Step 8: API and cloud models

Unsloth serves an OpenAI-compatible API, so anything that speaks OpenAI can point at it.

It also connects outward to OpenAI, Anthropic, Ollama, llama.cpp, and vLLM, letting you use cloud and local models in one interface with prompt caching preserved.

Known beta issue: tool access for connected cloud models can be inconsistent depending on the provider.

Step 9: Remote access (read this part properly)

By default Unsloth binds to 127.0.0.1 and is reachable only from your own machine. To reach it from your phone or another device:

powershell

# Recommended: HTTPS through a free Cloudflare tunnel
unsloth studio --secure -p 8888

powershell

# LAN only, trusted networks
unsloth studio -H 0.0.0.0 -p 8888

The security part. Server-side tools including web search and code execution run as your user and are enabled by default. Anyone who reaches your server with your API key can run code on your machine. Keep the key private, and pass --disable-tools if you’re exposing Unsloth anywhere public.

Two things Unsloth does well here: --secure fails closed, so if the tunnel can’t start, your raw port is never exposed. And the first time you publish a public URL with the auto-generated password still in place, it forces you to set a real one before the link goes live.

Step 10: Datasets and training

Data Recipes builds training datasets from PDF, CSV, JSON, and DOCX files, so you don’t need a prepared dataset to start.

Training supports LoRA, QLoRA, full fine-tuning, pretraining, and reinforcement learning (GRPO, DPO, FP8), running roughly 2× faster on about 70% less VRAM. Text, diffusion, text-to-speech, and embedding models are all supported. Export to GGUF, NVFP4, or FP8 when you’re done.

Troubleshooting

Generation feels slower than other apps. Web search, code execution, and tool-call healing all add time. Turn them off and speed should match any other llama.cpp app. If it’s still slow, open a GitHub issue.

GPU not detected. Confirm your driver is current. For Intel GPUs and other Vulkan-capable cards, reinstall with UNSLOTH_FORCE_VULKAN=1 set beforehand. Vulkan accelerates GGUF inference only; training still needs a supported PyTorch backend.

Existing models not showing. Set a custom model directory in settings.

Removing model files. Delete them from the bin icon in model search, or clear the Hugging Face cache at %USERPROFILE%\.cache\huggingface\hub\.

Uninstalling Studio. In PowerShell:

powershell

irm https://raw.githubusercontent.com/unslothai/unsloth/main/scripts/uninstall.ps1 | iex

To remove just the install directory and keep the shortcut for a reinstall:

powershell

Remove-Item -Recurse -Force "$HOME\.unsloth\studio"

Neither touches your downloaded models. To remove Unsloth Desktop, use Settings → Apps → Installed apps in Windows.

Why I replaced three tools with one

I was running Ollama as a local endpoint for agents, LM Studio for chatting and downloading models, and Open WebUI on top for search and document chat. Three installs, three configs, three update cycles.

Unsloth Desktop covers all three jobs and adds training, which none of them do. It’s dual licensed under Apache 2.0 for the core and AGPL-3.0 for the Studio UI, and the repo has around 70,700 stars.

Models you download from Hugging Face carry their own licenses. Check those before you use anything commercially.