What is Unsloth?

Unsloth is an open-source project for fine-tuning and running language models quickly and in little memory. It is best known for LoRA and QLoRA runs on a single GPU, and it publishes converted and quantized checkpoints on the Hugging Face Hub.

What it does

Unsloth makes fine-tuning cheaper in time and memory. It loads a model through its own optimized code and trains it with LoRA, QLoRA or full fine-tuning, on top of Hugging Face Transformers and TRL. It is best known for adapter runs that fit on one GPU, often in a notebook.

It comes as code (Unsloth Core) and as a desktop app and web UI. The project also publishes many model checkpoints on the Hugging Face Hub, converted or quantized so that they load directly.

When to use it

Unsloth fits when memory is tight and you want results fast: a quick experiment, a small or mid-size model, a LoRA or QLoRA adapter on one GPU.

Unsloth and Halo

The two meet in Halo's GPT-OSS cookbook, which trains from Unsloth's BF16 (16-bit float) checkpoints, unsloth/gpt-oss-20b-BF16 and unsloth/gpt-oss-120b-BF16. The native release stores its experts quantized, in MXFP4, and expert parallelism needs floating-point expert weights.

Halo is built for models that have to be split across GPUs or nodes. Reach for it for a mixture-of-experts model that needs expert parallelism, or for online reinforcement learning with rollout servers.

Train your own models

Post-train any Hugging Face model with Halo. Up to 4× faster than TRL