The coding models built for maintainability. Trained on real and synthetic data to write clean, disciplined code that respects your legacy and eliminates slop.
Read the researchLumen is optimized for the outcomes that matter: shipping better code, faster, with fewer headaches.
GPT-5.4
Gemini 3.1 Pro
GPT-5.5
Kimi K2.6
Lumen Outpost
Most AI models try to be everything to everyone. They write poetry, solve physics problems, and summarize novels.
Lumen does one thing: write exceptional software.
Our thesis is simple. A model trained exclusively on real production code, from real engineering teams, will always beat a generalist at the task developers actually need.
That is why every training decision, every optimization, every evaluation metric is grounded in a single objective: better code.
Generic reward signals produce generic results. Lumen was trained with a fundamentally different approach: we optimize for the characteristics that define truly great software.
Most models optimizes for getting to "it works." Lumen optimizes for "it works and you will thank yourself six months from now." We explicitly train against bloat: unnecessary abstractions, redundant logic, and over-engineered solutions that become tomorrow’s technical debt.
Too many AI interactions are exhausting and irritating. We train directly against this: not just getting to the right answer, but doing so with the right cadence, tone, and judgement. Clear updates, less filler, and better calls on when to act, explain, or push back.
Lumen is trained for the messy edges of real engineering: brittle tests, large diffs, ambiguous issue reports, and projects where the right answer requires reading before writing. The goal is not generic helpfulness; it is resilient software work.
Lumen isn't a single model. It's a system designed for every layer of the development workflow. From on-device utility to Sovereign reasoning, you use the right model for the job.
Capability
Utility tasks, codebase mapping, retrieval, indexing
Best for
Large-scale support workflows
Min GPUs
4 H100s
Our work has been independently validated across leading industry platforms and events.
cosine
Scout 
Download for Mac
Linux