What is Unsloth Desktop?

Unsloth Desktop is a free, open-source desktop application for Windows, macOS, and Linux that lets you run and fine-tune AI models locally on your own hardware. It supports 500+ models across text, vision, audio, and embeddings.

How do I connect Unsloth to Claude Code or Codex?

Load a model in Unsloth Desktop, open your project folder, then run 'unsloth start claude' (or 'unsloth start codex'). This routes your coding agent through your local model via Unsloth's OpenAI-compatible API.

What hardware does Unsloth support?

CPU, Apple Silicon (Metal), NVIDIA (RTX 30/40/50, Blackwell, DGX), AMD (ROCm), Intel (Vulkan), and multi-GPU setups. Training requires a supported PyTorch or MLX backend; inference also works via CPU and Vulkan.

How much faster is Unsloth training?

Unsloth claims 2x faster training with 70% less VRAM compared to standard methods, with no accuracy loss. This is achieved through kernel optimizations and memory-efficient training techniques.

Yes, Unsloth Desktop is completely free and open-source. The code is on GitHub (unslothai/unsloth) under the Apache 2.0 license.

Can I use Unsloth models remotely?

Yes, Unsloth includes Cloudflare tunnel integration for secure HTTPS access to your local models from anywhere. Run 'unsloth studio --secure' to enable it.

What are the best use cases for local model training?

Domain-specific fine-tuning (legal, medical, finance), proprietary data that can't leave your infrastructure, custom coding assistants trained on your codebase, cost reduction for high-volume inference, and privacy-sensitive applications.

How does Unsloth compare to OpenPipe ART for agent training?

Unsloth focuses on efficient supervised fine-tuning and LoRA/QLoRA workflows via a GUI. OpenPipe ART specializes in reinforcement learning for multi-step agents. They're complementary — you might fine-tune a base model with Unsloth, then train it for agentic tasks with ART.

Unsloth Desktop: Train and Run LLMs Locally, Then Connect Them to Claude Code

By Prahlad Menon Published 2026-08-17 5 min read

Local AI just got real.

Unsloth shipped their desktop app this week — and it’s not just another inference tool. This is the first mainstream application that does both training and running of LLMs locally, wrapped in a GUI instead of a Python environment.

The game-changer? One command connects your local model to Claude Code, Codex, or OpenClaw.

What Unsloth Desktop Actually Does

Three core capabilities that matter:

1. Run 500+ Models Locally

Text, vision, audio, embeddings — Unsloth supports them all:

LLMs: Qwen3.8, DeepSeek-V4, Gemma 4, Kimi K3, MiniMax-H3, Llama 3.x
Vision: Gemma 4 E2B, Qwen3.5 Vision, Ministral 3 VL
Audio: Orpheus-TTS
Embeddings: EmbeddingGemma

Models download and run with a click. GGUF format means efficient quantization for consumer hardware.

2. Fine-Tune on Your Hardware

This is where Unsloth shines. The numbers:

2x faster training than standard methods
70% less VRAM usage
No accuracy loss

Supports the full training stack:

LoRA and QLoRA (the efficient way to fine-tune large models)
Full fine-tuning (when you have the VRAM)
GRPO/DPO reinforcement learning
Pre-training from scratch

All from a GUI. No Python required.

3. Connect to Coding Agents

Here’s the killer feature for developers:

unsloth start claude

That’s it. Unsloth serves an OpenAI-compatible API, Claude Code connects to it, and suddenly your coding agent is running through your local model instead of hitting the cloud.

Supported agents:

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

You can even run it as a subagent — Claude Code keeps its cloud model for main reasoning but calls your local model for specific tasks:

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

Hardware Support

Unsloth runs on basically everything:

Hardware	Inference	Training
CPU	✅	✅ (slow)
Apple Silicon	✅ (Metal)	✅ (MLX)
NVIDIA RTX 30/40/50	✅	✅
NVIDIA Blackwell/DGX	✅	✅
AMD (ROCm)	✅	✅
Intel/AMD (Vulkan)	✅	❌
Multi-GPU	✅	✅

The Vulkan backend is particularly interesting — it enables GPU-accelerated GGUF inference on hardware that doesn’t have CUDA or ROCm support, including some Intel integrated GPUs.

Installation

Desktop App (recommended):

Or install via terminal:

# Linux/macOS
curl -fsSL https://unsloth.ai/install.sh | sh

# Windows PowerShell
irm https://unsloth.ai/install.ps1 | iex

Python package (for code-based workflows):

uv pip install unsloth --torch-backend=auto

The Local AI Stack is Complete

Here’s what’s happened:

Inference tools (Ollama, LM Studio, llama.cpp) made running models locally easy
Fine-tuning libraries (Unsloth Core, PEFT, TRL) made training possible
Desktop apps (Unsloth Desktop) made both accessible without code

And now with the Claude Code/Codex integration, local models slot directly into developer workflows.

The missing piece was always “okay I have a local model, now what?” Unsloth’s answer: wire it into the tools you already use.

Real Use Cases for Local Training

Why would you train locally instead of using API-based fine-tuning services? Here are the scenarios where Unsloth shines:

1. Domain-Specific Models (Legal, Medical, Finance)

You have proprietary documents — contracts, medical records, financial reports — that cannot leave your infrastructure. Fine-tune a model on your data, deploy it locally, and your sensitive information never touches the cloud.

Example: A law firm trains Qwen 14B on their case history. The model learns their document formats, legal citations, and writing style. It runs on their server, accessible only internally.

2. Custom Coding Assistants

Train a model on your codebase. It learns your architecture patterns, naming conventions, internal APIs, and coding style. Then connect it to Claude Code as a subagent for repo-specific tasks.

# Train on your repo's code
unsloth train --data ./my-codebase --model qwen3-8b

# Connect to Claude Code
unsloth start claude --as-subagent --model ./checkpoints/my-codebase-lora

3. High-Volume Inference Cost Reduction

If you’re making thousands of API calls daily, the math changes. A $2,000 GPU pays for itself in months compared to API costs. Train once, run forever.

4. Privacy-Sensitive Applications

Healthcare chatbots, HR assistants, customer support for regulated industries — any application where conversation data must stay on-premises.

5. Reinforcement Learning for Agents

Unsloth supports GRPO and DPO for training agents with reinforcement learning. If you’ve read our coverage of OpenPipe’s ART framework, Unsloth provides a complementary approach: efficient base model fine-tuning that you can then enhance with RL techniques.

The workflow: Fine-tune a base model with Unsloth (supervised learning on your data) → Train agentic behaviors with ART (reinforcement learning on task trajectories) → Deploy locally or export.

Why This Matters

For developers: Your coding agent can now run on your GPU, with your fine-tuned model, keeping your code private. The cost is electricity instead of API calls.

For companies: Train domain-specific models on proprietary data without it leaving your infrastructure. Deploy the result locally or export to GGUF/FP8 for production.

For researchers: One-click access to training workflows that used to require hours of environment setup. Export to any format.

For hardware buyers: Finally a reason to use that RTX 4090 for more than gaming. Unsloth’s VRAM efficiency means 70B models fit where 30B used to struggle.

The Competition

Tool	Run Models	Train Models	Desktop App	Agent Integration
Unsloth	✅	✅	✅	✅
Ollama	✅	❌	❌ (CLI)	Partial
LM Studio	✅	❌	✅	✅
llama.cpp	✅	❌	❌	❌
Axolotl	❌	✅	❌	❌

Unsloth is the first to check all four boxes.

What’s Next

The MCP (Model Context Protocol) integration is already in PR — letting compatible clients manage models, training, checkpoints, and exports directly. The local AI stack is converging on standard protocols.

Remote access via Cloudflare tunnels is also built in (unsloth studio --secure), so you can serve your local models securely without port forwarding.

Links: