What This Is
Anthropic has a model architecture called Mythos. It has never disclosed any technical details about it. Kye Gomez — 22 years old, founder of the a gentic framework Swarms — assembled a set of public papers and industry speculation, pieced together what he believes approximates the Mythos architecture, and released it as an open-source project under the name OpenMythos.
The core idea behind the architecture is called the Recurrent-Depth Transformer (RDT): the same set of model parameters runs multiple computation passes, activating different "expert modules" on each pass, rather than stacking hundreds of layers with distinct parameters as mainstream approaches do. Concretely: the same weights execute up to 16 passes, each taking a different expert pathway. The entire reasoning process unfolds inside an internal vector space with no intermediate steps exposed — only a final answer is returned.
On parameter efficiency, a paper from UCSD and Together AI shows that a 770M-parameter RDT model matches a 1.3B-parameter standard Transformer — nearly half the parameter count, equivalent performance. More significantly, on generalization: when tested on " knowledge combinations not seen during training," the recurrent architecture still produces correct answers where standard models fail outright.
Industry View
Those who support this direction argue that the bottleneck for large models today is no longer "how many facts were memorized" but "whether known facts can be combined to answer novel questions." Running a recurrent architecture for additional passes at inference time appears to confer this combinatorial ability essentially for free, without expanding training scale. If that judgment holds, the competitive axis of the AI industry's next phase shifts from "train bigger models" to "make existing models reason deeper at inference time."
The counterarguments deserve equal s eriousness. First, OpenMythos is fundamentally a speculative reconstruction — Anthropic has never confirmed that Mythos uses this architecture, and Gomez himself acknowled ges it is "an integration of mainstream guesses." Second, the 770M-versus-1.3B experiment is small -scale; there is no evidence yet that the findings repl icate at larger parameter counts. Third, the stability of recurrent inference (keeping each computation pass from diverging) currently depends on specific injection mechanisms whose engineering reliability has not been validated at scale. Academic interest in a direction is one thing; production-grade deployment is another.
Impact on Regular People
For enterprise IT: If the "fewer parameters, more inference passes" approach is validated by the mainstream, the hardware barrier to enterprise AI deployment may fall. Achieving equivalent capability with smaller models means the cost calc ulus for on-premises private deployment gets recalculated from scratch.
For individual careers: This class of technical development will not change how individuals use AI tools in the short term. But it signals that AI's capability ceiling on "complex reasoning" and "knowledge synthesis" is being pushed further out — and the pace at which " judgment work" gets displaced by AI may arrive faster than most expect.
For the consumer market: Recurrent inference architectures are naturally suited to "think slowly, deliver an answer at the end" scenarios rather than real-time conversation. Consumer-facing products will feel no change in the short term, but in the medium term we may see a more clearly differentiated product landscape: a "deep reasoning mode" and a " fast response mode" as distinct, labeled tiers.