Someone reverse-engineered Anthropic's rumored Claude Mythos architecture from public research papers and shipping hints—OpenMythos by @kyegomez is now live on GitHub as a working PyTorch implementation.

Architectural breakdown:

• Recurrent-Depth Transformer: Instead of stacking N unique layers, it loops a smaller set of recurrent blocks. Think of it as vertical depth replaced by horizontal iteration.

• Sparse MoE with ~5% activation: Total param count is in storage, but only a tiny fraction fires per forward pass. Efficient at scale.

• Loop-index positional embeddings: Each recurrence step gets its own positional signal, treating iterations as computational phases rather than token positions.

• Adaptive Computation Time (ACT) halting: The model dynamically decides when to stop "thinking" per token. No fixed depth—it halts when confidence threshold is met.

• Continuous latent thoughts: Internal state carries over across iterations, enabling breadth-first search-style reasoning instead of purely autoregressive left-to-right.

This isn't confirmed to be Claude Mythos 1:1, but it's a fully cited, runnable hypothesis. Every design choice maps back to actual papers. Whether Anthropic uses this exact stack or not, OpenMythos is a solid reference implementation for anyone exploring recurrent transformers, dynamic compute, and next-gen reasoning architectures.

Code is public. Worth pulling and profiling if you're into model internals.