companies

Unsloth AI

Libreria open source per il fine-tuning rapido ed efficiente nell’uso della memoria di LLM a pesi aperti su GPU consumer e su singole GPU.

In inglese

Unsloth AI was founded in 2023 by brothers Daniel Han and Michael Han, who built the project after Daniel's work at NVIDIA optimizing algorithms like t-SNE and later hunting down bugs in open-weight models such as Gemma, Llama, and Phi. The company went through Y Combinator and has stayed a lean, engineering-heavy team focused almost entirely on making fine-tuning faster and cheaper rather than building a broad platform. Unsloth's core product is an open-source Python library that rewrites the training kernels used for LoRA and QLoRA fine-tuning, cutting VRAM usage enough to fine-tune a 7-billion-parameter model on a single 24GB consumer GPU. It supports hundreds of open-weight model families and ships free notebooks for Google Colab and Kaggle so anyone can fine-tune without paying for compute beyond what those platforms already provide. Paid Pro and Enterprise tiers add faster training plus multi-GPU and multi-node support for teams that outgrow a single card. The library has become a default reference point in open-source LLM communities on Reddit and Hugging Face, with millions of monthly downloads and thousands of community-trained models built on top of it.

Founded

2023

Headquarters

San Francisco, United States

Sector

infrastructure