Unsloth

What is Unsloth?

Unsloth is an open-source framework for running and training LLMs on your own hardware. It supports inference and training for 500+ models, including vision, text-to-speech, embedding and reinforcement learning models, and runs on macOS, Linux and Windows with NVIDIA, AMD, Intel and CPU.

Key capabilities of Unsloth

• Faster fine-tuning: the docs claim 2x faster training with 70% less VRAM.
• Training methods: LoRA, QLoRA, full fine-tuning, pretraining, RL, GRPO, DPO and FP8.
• Local inference: run LLMs, diffusion, embedding and audio models with an OpenAI-compatible API.
• Unsloth Studio: browser-based UI for running and training models.
• Export: to GGUF, NVFP4 and FP8.
• Data Recipes: build datasets from PDFs, CSVs and DOCX files.

How Unsloth works

You install Unsloth natively or through Docker, Homebrew or WinGet, or use the Studio desktop apps. You then load a model, fine-tune it with your chosen method, and export it in a format such as GGUF to run elsewhere. Local agents can be run, including with Claude Code, Codex and MCP.

Who uses Unsloth?

ML engineers, researchers and hobbyists who want to adapt open models on limited GPU memory. It targets people who run models on their own machines rather than through a hosted API.

Unsloth pricing

Unsloth is an open-source framework, and its documentation does not state a price. Costs come from the hardware or cloud GPUs you use.

Unsloth alternatives

Lamini offers an enterprise platform for tuning models, so it is a managed service rather than a local framework. Arcee AI builds small language models and the tooling around them. Nous Research releases open models and is a model developer rather than a training toolkit.