OpenMythos rebuilds Claude Mythos in PyTorch

Interesting more as a research harness than a claim of fidelity: it turns a speculative architecture into something you can inspect, run, and stress-test. The real value is making loop depth, halting, and expert routing measurable inside a concrete PyTorch codebase.

• –It translates Claude Mythos speculation into executable architecture, which is useful even if the reverse engineering is only approximate.
• –The recurrent block plus adaptive looping is the core idea: same weights, more passes, deeper internal refinement.
• –Sparse MoE with shared experts fits current scaling instincts, but the architecture still reads as a hypothesis until trained weights or model cards exist.
• –For researchers, this is a practical sandbox for testing whether latent iterative reasoning actually beats standard feed-forward depth.