Intelligence anyone can build. And everyone can verify.
Our deepest thanks to everyone at ApexSI, and to you for visiting.
We are a company researching all the way to AGI, so that all of humanity can share in a common intelligence. ApexSI builds language models in the open: every model ships with its weights, its training code, its training logs and the research that explains how it works.
Thank you for joining us on this journey.
Models3 released, 1 in trainingApex-2, Apex-2-Base and Apex-1-DPO on Hugging Face.
Apex-321B MoE on one GH2003.36B active per token, pretraining on real data.
CodeThe full training stackModel, optimizer, offload and FP8, on GitHub.
LogsEvery step, on the recordLoss, throughput and MFU, per model.
ResearchEvery claim, measuredHow Apex-3 trains, including what failed.
Models