Meet OpenMythos: An Open-Source PyTorch Reconstruction of Claude Mythos Where 770M Parameters Match a 1.3B Transformer

OpenMythos is an open-source PyTorch project that proposes Claude Mythos may use a recurrent-depth transformer design.

The hypothesized architecture combines a Prelude, a looped recurrent block, and a Coda, with MoE routing and multi-latent attention.

The project argues the model may reason in continuous latent space instead of generating intermediate text step by step.

If true, this would suggest iterative computation can improve reasoning efficiency and capability without a proportional increase in parameters.