Skip to content

Unsloth

Fine-tune open models faster with far less memory

160

Overview

Unsloth is an open-source desktop application that lets you train, run, and serve open models on your own hardware, aimed at developers and researchers who want local control over their AI workflows.

Local Image and Video Generation

Unsloth Desktop includes tools for creating images with models like Qwen-Image-2.1, MiniMax-H3, FLUX, and Z-Image, along with LoRA adapters. Video generation is supported through Wan and LTX. The app also handles image editing tasks such as transform, inpaint, extend, upscale, reference, and edit, all performed locally. On an NVIDIA B200, MiniMax-H3 generated a 960×544, 124-frame video in 13 seconds at 8 steps, a notable improvement over previous methods that took 70 seconds or more.

Connecting Agents to Local Models

Developers can connect agents like Claude Code and Codex to local models using the unsloth start command. The workflow involves starting Unsloth, loading a model, opening a project folder, and then running unsloth start claude. Models can be swapped within Unsloth without altering the agent workflow. Unsloth also provides an OpenAI-compatible API, allowing existing applications, scripts, and SDKs to connect to local models through a familiar interface.

Private Web Search and Deep Research

Unsloth Desktop includes unlimited, private, and free web search. Users can search while the model is thinking or use Deep Research to find sources and generate detailed reports with citations. This feature keeps research activities within the local environment, avoiding external data exposure.

Code Execution and Tool Calling

Models running in Unsloth can execute Bash and Python in a secure sandbox, enabling them to run code, test results, and complete tasks locally. The platform claims up to 50% more accurate tool-calling with self-healing tool calls that detect, repair, and retry failures automatically. This makes it suitable for agentic workflows that require reliable code execution.

Model Hub and Remote Access

Unsloth supports nearly every open model, with Day Zero support for releases like Qwen3.8, NVIDIA, GLM, and Gemma. A built-in model hub helps discover, manage, and download the right quantization for your device. Users can also serve local or Colab models over HTTPS through a free Cloudflare tunnel, allowing access from a phone, laptop, or other devices.

Training Performance and Plans

Unsloth offers a free open-source version that supports Mistral, Gemma, and Llama models with 4-bit and 16-bit LoRA, though multi-GPU support is listed as coming soon. The Pro plan provides 2.5x faster training and 20% less VRAM usage, with enhanced multi-GPU support for up to 8 GPUs. The Enterprise plan adds 32x faster training, up to 30% accuracy boost, 5x faster inference, full training, multi-node support, and customer support. Pricing for Pro and Enterprise requires contacting Unsloth, and specific costs are to verify.

Core features

  • Open-source desktop UI for training and running open models locally
  • Image generation with Qwen-Image-2.1, MiniMax-H3, FLUX, Z-Image, and LoRA adapters
  • Video generation with Wan and LTX
  • Image editing tools: transform, inpaint, extend, upscale, reference, and edit
  • Agent connectivity via `unsloth start` command for Claude Code, Codex, and others
  • OpenAI-compatible API for existing apps, scripts, and SDKs
  • Built-in model hub for discovering, managing, and downloading quantizations
  • Private, unlimited web search and Deep Research with citations

Use cases

  1. 1Training and fine-tuning open models on local GPUs
  2. 2Generating images and video locally with popular open models
  3. 3Connecting coding agents like Claude Code or Codex to local models
  4. 4Serving local or Colab models over HTTPS via Cloudflare tunnel for remote access
  5. 5Running sandboxed Bash and Python code for tool-calling and testing
  6. 6Conducting private web research with Deep Research and citations

Strengths and things to check

Strengths

  • Runs entirely on local hardware, keeping data private
  • Supports a wide range of open models with Day Zero releases
  • OpenAI-compatible API simplifies integration with existing tools
  • Free web search and Deep Research built into the desktop app
  • Self-healing tool calls improve accuracy by up to 50%

Things to check

  • Requires capable local hardware, especially for video generation and large models
  • Multi-GPU support is limited in the free version and coming soon
  • Advanced training features like 2.5x speed and 20% VRAM reduction require a Pro plan
  • Enterprise features such as multi-node support and 30x speed are only available in the Enterprise plan

Pricing

Unsloth offers a free open-source version with support for Mistral, Gemma, and Llama models, 4-bit and 16-bit LoRA, and single-GPU training. The Pro plan adds 2.5x faster training, 20% less VRAM, and multi-GPU support up to 8 GPUs. The Enterprise plan includes 32x faster training, up to 30% accuracy boost, 5x faster inference, full training, multi-node support, and customer support. Pricing for Pro and Enterprise requires contacting Unsloth.

Basic information

Website
unsloth.ai
Added
Oct 8, 2026
Platforms
Web
Chinese & English
Not marked
API
Not marked
Self-hosting
Not marked
Best for
Developers, researchers, and AI practitioners who want to train, run, and connect open models locally without relying on cloud services.
Information updated
Oct 16, 2026
Website status
Checking
Corrections · Privacy · Data requests?
Contact us

Features, pricing and availability may change. Always verify details on the official website. Maintained by steven · Last updated Oct 16, 2026 →

Related tools

All categories