About us

Axiomic Labs trains language models small enough to run on a laptop CPU or a phone. Our research question is how far careful data design can take a model when its parameter count is fixed and small.

GPT-X3-150M, our latest model, scores higher than Hugging Face's SmolLM2-135M on the Open SLM Leaderboard after training on 75 billion tokens, against SmolLM2's 2 trillion. We trained it on two consumer RTX 3080 Ti graphics cards over about 330 hours.

How we work

All of our language models are pretrained from random weights. We don't fine-tune other labs' models. Models, datasets and benchmarks are published on Hugging Face under permissive licences, and we score our models on the Open SLM Leaderboard with the same method we use for everyone else's.

Team

Dan Prichard

Founder

Ethan Le Sage

Operations & Safety

Michael Morgan

Research & Development