kyegomez / OpenMythos

What is OpenMythos?

OpenMythos delivers a pip-installable Python package for building recurrent-depth transformers, theorized to power Claude's advanced reasoning. Drawing from available research literature on theoretical computer science GitHub threads and papers, it reconstructs the Mythos architecture from first principles: a prelude of standard layers, a looped recurrent block for depth-variable thinking, and a coda for output refinement. Users get a configurable model that runs forward passes, generates text, and exposes stability metrics like spectral radius—all in pure Torch.

Why is it gaining traction?

With 411 stars, it stands out by enabling inference-time depth scaling: crank up loops for harder reasoning tasks without retraining or exploding parameters, mimicking Claude's systematic generalization on novel problems. Switchable attention modes and sparse MoE feed-forwards deliver efficient compute-adaptive depth, letting devs probe looped transformer hypotheses hands-free. The hook? Instant prototyping of research ideas from Twitter papers, no from-scratch coding needed.

Who should use this?

AI researchers dissecting theoretical physics-inspired architectures on GitHub, or ML engineers testing looped models for multi-hop math, planning, or code reasoning. Ideal for PhD students replicating Parcae scaling laws, or teams exploring parameter-efficient alternatives to dense giants like Llama.

Verdict

Grab it for theory-driven experiments—docs and tests are thorough for an alpha project—but temper expectations with its 0.9% credibility score and modest stars; it's speculative reconstruction, not production-ready Claude clone. Solid starting point for recurrent research.

(198 words)