We ran a 9B model against Anthropic's Mythos on Firefox. See the early results (shipitclean.com)

In an intriguing contest of computational capability, a 9-billion parameter model, Roasty, has been pitted against Anthropic's vast 10-trillion parameter model, Mythos, in analyzing Mozilla Firefox's code for security vulnerabilities. As of April 28, 2026, Roasty has processed 43% of its scan, uncovering 377 potential security issues compared to Mythos's 271 findings. Notably, Roasty utilizes a specialized team of reviewers to flag vulnerabilities across different dimensions, striving for accuracy over quantity. The ongoing results suggest that even a smaller model may match or exceed the findings of a significantly larger one, challenging the assumption that size alone dictates effectiveness in security research. The implications for the AI/ML community are profound: if Roasty continues to perform well, it may democratize access to effective security auditing tools, making high-level analysis feasible without the need for extensive computational resources. This could shift the focus from sheer model size to architecture and specialized techniques, as Roasty's innovative design incorporates a deterministic retrieval engine, Atlas, which preserves accuracy and context when scanning vast codebases. Should Roasty achieve results comparable to Mythos, it may redefine the benchmarks for code vulnerability assessments, potentially broadening participation in secure coding practices and reducing barriers for developers and security researchers alike.