IOTA tackles one of the most challenging and important issues of our time - the training of massive AI models. Our philosophy is to build in public and iterate quickly. This means both transparent experiments, and permissionless participation.
IOTA coordinates global compute to create an indistinguishable compute fabric from distributed, hetereogenous nodes. This unlocks truly collaborative open models, and a new paradigm in cost-efficient training.
Select individual experiments or view all at once. Loss curves for highlighted experiments. Select runs below to compare.
Median loss at 50M tokens. Lower is better.
Tokens per replica second (TPRS). Higher is better.
Train at Home hardware breakdown across contributing compute nodes.