Unsloth

Freemium

Unsloth is a desktop app for training and running AI models on your own hardware. You keep your data local while you work with text, image, video, and audio models.

You can fine-tune models through a no-code interface or connect coding agents like Claude Code and Codex. The OpenAI-compatible API lets existing tools talk to your local models.

This is a Coding Assistants tool for people who need local control over model training and inference.

Open it if you want to train or run models without cloud services or sending data out. Skip it if you want polished consumer software or need multi-GPU training without paying.

It fits when you need to keep data local or avoid cloud costs.

• Users wanting local LLM training with LoRA/QLoRA: You get the full training workflow in one GUI. No command-line tools needed.
• Developers needing OpenAI-compatible local API: You plug your own models into existing apps without rewriting code.
• AI developers wanting local training and inference: You keep everything on your machine with no telemetry.
• Privacy-conscious users requiring no telemetry: Your data never leaves your hardware.
• Practitioners fine-tuning open-weight LLMs on limited GPU memory: You fit larger models on consumer GPUs with less VRAM.
• Solo researchers and small companies on single-GPU setups: You skip cloud infrastructure and keep costs low.
• Anyone training reasoning models with limited VRAM: You run GRPO training on hardware like an RTX 4090.
• Users who want to train or fine-tune models without command-line tools: You annotate data, train, and infer through one interface.

These parts stick out once you start using it.

• Complete training workflow in one place: You prepare data, train, and test results without switching tools. The GUI replaces scattered scripts and notebooks.
• OpenAI-compatible local API: Your existing tools and agents connect directly. You do not rebuild workflows to go local.
• Multiple export formats: You save to GGUF, vLLM, safetensors, or merged weights. One tool handles deployment prep.
• Auto-discovery of existing models: It finds models you already downloaded from other runners. You skip reimporting and folder hunting.
• Deep per-model controls: You adjust KV cache, speculative decoding, and VRAM budgets per model. The granularity matches or beats alternatives.
• Tool call permissions layer: You set Approve, Always Approve, or Deny for each tool. Other local runners do not have this.

Skip it when these limits block your work.

• Still in beta: Updates come fast and can break things. You pin versions and test before trusting it.
• NVIDIA-focused: AMD and Apple Silicon support lags or is missing. You need an NVIDIA GPU for full training.
• Single-GPU only in free version: Multi-GPU training needs a Pro subscription. You pay to scale beyond one card.
• Console window stays open: The interface runs in a browser tab, not a native desktop app. You see a terminal window in the background.
• Patchy documentation: Key features lack complete docs. You reverse engineer error messages or wait for community help.