Cosine

Cosine describes itself as a sovereign AI lab. It trains AI models and specialised coding agents for organisations that need frontier capability inside secure environments. Its model family is called Lumen, with versions named Lumen Sovereign, Lumen Outpost and Lumen Scout. Cosine is based in the United Kingdom and has published an open letter on sovereign AI for the UK, supported by a coalition of executives from critical national infrastructure.

Specialists over generalists

Cosine's thesis is that specialists outperform generalists. Most AI models try to do everything, from poetry to physics. Lumen does one thing: write software. Cosine says it is trained on real production code from real engineering teams, and every training decision and evaluation metric is aimed at better code. It emphasises less AI slop, meaning output optimised for maintainability, readability and production quality rather than bloated code.

Benchmarks

Cosine publishes its own benchmarks. Niche-Bench measures performance across niche and legacy programming languages where standard coding benchmarks fall short. Vibe-Bench measures how reliably models turn high-level intent into working code. Slop-Bench measures whether code is maintainable, precise and free from common low-quality patterns. A cost per successful task metric compares the effective cost of completing a task, rather than the price of a single model call. On its Niche-Bench chart, Lumen Outpost scores 53.9 percent against figures between 42.4 and 48.3 percent for other named models. These are vendor-run results.

Surfaces and solutions

Lumen is available through a CLI, a desktop app and the cloud. Solutions target sovereign AI, air-gapped deployments, cybersecurity and niche languages. A red-team application and training are listed as coming soon.

Pricing

Cosine has a pricing page and a Try our models option, and organisations can contact the team for secure deployments.

Who it suits

Cosine suits government, defence, critical infrastructure and regulated enterprises that cannot send code to public AI services, and teams maintaining legacy codebases.

Things to consider

Validate benchmark claims on your own code. Review all generated code and keep normal testing and security review in place.

For air-gapped use, plan how models will be updated and supported over time, and run a pilot on a representative codebase before committing.