The Sovereign Frontier Model
Lumen Sovereign is Cosine’s planned sovereign model. It is being developed for organisations that need frontier-capable AI while keeping control over where models run, how data is governed, and how systems are deployed.
In coalition with
Leading the UK sovereign AI story
Co-designed with major UK institutions
Coalition members will help define the use cases, security requirements, and governance standards Lumen Sovereign must meet.
Training on UK sovereign compute
Developed using compute awarded through the UK Government’s Sovereign AI programme, with training on Isambard-AI.
Frontier capability for sovereign AI
Lumen Sovereign is being built for complex work in regulated environments, combining frontier-level ambition with UK-controlled training, governance, and deployment.
2026
Q2
Baseline & Architecture Validation
Test MoE scaling pilots, lead configuration expected to be 192–256 experts / top-6 routing, with promotion based on loss, SWE evals, long-context checks, router & expert stability, and performance.
Q3
Model Enlargement
Run the main expert-expansion CPT phase to produce Lumen Sovereign 256K, a 1T-parameter intermediate checkpoint with 49.2B active parameters. Partners receive progress updates and previews of capabilities.
Q3
Agent Post-Training
Run supervised fine-tuning on high-quality agentic demonstrations tailored to coalition and public RFI workflows (across generalist reasoning and coding), followed by asynchronous reinforcement learning on long-horizon executable coding tasks.
Q4
Alpha Access
Coalition members gain early access to the model at key checkpoints, alongside enterprise applications for internal evaluation.
2027
Q1
Commercial Launch
Full deployment, available on cloud, VPC, or fully air-gapped on your own GPUs.
Built for the environments where AI cannot leave the perimeter
Lumen Sovereign is being designed for the world’s most secure industries, where sensitive code, data, and operational context must stay inside controlled infrastructure.
Runs inside controlled customer environments
Supports private and dedicated deployment models
Compatible with fully air-gapped systems
Keeps sensitive code and data inside approved infrastructure
Designed for governance, auditability, and security review
Gives regulated teams a path to frontier-capable AI without relying on foreign-hosted models
The frontier model in the Lumen family
Scout
Utility model for codebase mapping, retrieval, indexing, and high-volume support workflows. Use when you need low-cost intelligence at scale.
Outpost
Production coding model for enterprise AI-assisted software engineering across teams. Use when you need strong coding performance, maintainability, and practical cost efficiency.
Sovereign
Frontier sovereign model for architecture, deep reasoning, complex transformation, and controlled deployment environments. Use when performance and control matter above all else.