Cosine is an enterprise AI software-engineering platform ā now positioned as a sovereign AI lab ā that trains specialized coding models and deploys autonomous coding agents inside secure environments, up to and including fully air-gapped installations.
Cosine’s Lumen model family is post-trained exclusively on real production code, with a single objective: writing maintainable, production-quality software rather than generic text. Its agents work from a terminal-native interface or a shared cloud workspace, plan large changes, run multiple coding tasks in parallel, and call external tools via MCP. Deployment ranges from a managed public cloud to a dedicated single tenant or a fully air-gapped environment inside your own perimeter, and the models are specifically tuned for complex architectures, legacy infrastructure, and niche languages such as COBOL, Fortran, Verilog, Rust, and complex SQL.
Cosine emerged from the team behind Genie, an autonomous software engineer that posted state-of-the-art results on the SWE-bench benchmark, and it has since focused that agent technology on the enterprise: long-horizon tasks, sprawling legacy systems, and security postures where models must run entirely inside the customer’s own infrastructure. Its Scout model is post-trained from Devstral 123B and its Outpost model from Kimi K2.6, reflecting a strategy of specializing strong open foundations exclusively for software engineering.
Pricing page not published; contact sales. Cosine is sold as an enterprise platform with deployment-specific agreements ā managed cloud, single-tenant, or air-gapped ā rather than self-serve per-seat plans.
Cosine is built for regulated enterprises, financial institutions, and organizations with complex legacy estates that need frontier coding capability without their code ever leaving a controlled security perimeter.