OpenMythos is an open-source project that proposes Mythos as a Recurrent-Depth (looped) Transformer architecture. Anthropic describes Mythos as its most capable model to date, positioned a tier above Opus, and OpenMythos attracted more than 10,000 GitHub stars within weeks of release. Mozilla testing reported Mythos found 271 vulnerabilities in Firefox, and Mythos is held under Project Glasswing, a coalition of about 40 partners that includes Microsoft, Apple, Amazon, and the NSA.
OpenMythos defines a family of model variants that range from 1 billion to 1 trillion parameters. The repository README points to a 3 billion-parameter training script on FineWeb-Edu and specifies a Chinchilla-adjusted training target of 30 billion tokens. The project notes that training the larger variants would require substantial compute on H100 GPUs, with estimated costs in the hundreds of thousands of dollars. The repository contains training scripts and scaling targets but does not include completed trained models, and the project states that nobody has completed the proposed training.
To reduce memory demands and support very large parameter counts, OpenMythos incorporates DeepSeek’s Multi-Latent Attention to compress memory and employs a Mixture-of-Experts configuration to handle breadth across domains. The repo presents these mechanisms as part of its implementation strategy across the declared parameter range and links them to the provided training scripts and token targets. Those design elements are presented in the repository rather than as pretrained weights, so the codebase offers architectures and recipes rather than an executed model. Because the training work has not been completed, none of the defined 1 billion-to-1 trillion parameter variants have published weights.
The Parcae model, developed by the University of California San Diego and Together AI, boasts 770 million parameters. It aims to match the performance quality of a 1.3 billion fixed-depth transformer. The accompanying research paper discusses methods to address instability issues that often arise in looped models, proposing predictable scaling laws related to the depth of loops as a solution.
OpenMythos posits that its Mythos model operates as a Recurrent-Depth Transformer, commonly referred to as a looped transformer. This architecture is central to the OpenMythos hypothesis. Despite its intrigue, the public cannot access Mythos directly, as it remains a theoretical model without readily available execution parameters or weights.
Mythos was used in security testing that included finding 271 vulnerabilities in Firefox during Mozilla testing. The model also completed a 32-step corporate network attack simulation. Mythos is maintained within Project Glasswing, described as a coalition of about 40 partners. Partners in Project Glasswing include Microsoft, Apple, Amazon, and the NSA.
The above details summarize the security testing outcomes and the partner structure associated with Mythos. These points are presented in the project’s reporting.
OpenMythos is a speculative open-source project that models the Mythos AI architecture and has attracted notable community attention since its release. Mythos is maintained within Project Glasswing, a collaborative coalition of about 40 partners that includes Microsoft, Apple, Amazon, and the NSA, and the OpenMythos repository sets out architectural and scaling proposals related to a looped (Recurrent-Depth) transformer design and associated supporting mechanisms.
