Unsloth AI

Up to 30x faster training than FA2 is the headline claim for Unsloth AI, an open-source tool for fine-tuning and reinforcement learning on large language models.

Unsloth AI also cuts memory use by up to 90% compared with FA2, which is what lets you train on smaller GPUs.

It runs on NVIDIA, AMD and Intel hardware and supports a wide range of LLMs, so you are not locked into one model family or one vendor.

The project describes itself as beginner-friendly, and the core version is free.

The practical payoff is turnaround time.

Custom model training that might take 30 days can drop to around 24 hours, according to the project's own example.

Faster inference is included too, and the team says even faster inference is still in the works.

It is also pitched as more energy-efficient than standard training.

Multi-GPU support in the free version is listed as coming soon, so large multi-card jobs are not covered yet.

Pro and Enterprise plans exist, but you have to contact the company to get pricing.