Cosine has formed a coalition with leading UK institutions to build Lumen Sovereign: Britain's first fully sovereign frontier AI for critical sectors.
Today, Cosine announces the formation of a landmark industry coalition to co-design Lumen Sovereign, Britain’s first fully sovereign frontier AI model.
With input from some of the UK’s largest and most critical institutions – including BAE Systems, Babcock International Group, BT, Era4, HSBC, Leonardo, Lloyds Banking Group, LSEG, NatWest Group, PwC, Telefónica Tech UK&I, Thales UK, and The Alan Turing Institute – Cosine is building an AI capability that operates with zero dependence on foreign infrastructure.
Backed by the UK Government’s £500m Sovereign AI programme, Lumen Sovereign will be trained entirely on UK soil using Isambard-AI, one of Europe’s most powerful supercomputers.
Breaking the dependency risk in AI
For regulated sectors, AI has evolved from a productivity tool into a critical strategic dependency. Until now, the UK’s defence, financial, and public sector institutions have faced a difficult choice: either fall behind in AI adoption or rely on models trained and operated overseas, which can ultimately create unacceptable operational and sovereignty risks.
A sovereign LLM provides strategic advantages beyond avoiding dependency, including the ability to operate in high-security, air-gapped environments for complete domestic data control. This assurance-by-design is often non-negotiable in regulated sectors, ensuring alignment with specific governance standards and UK/Europe-specific regulatory requirements. Furthermore, it offers a competitive advantage by providing crucial cost protection against vendor lock-in and price escalations from expensive foreign-built APIs.
By signing a memorandum of understanding, Cosine and the coalition of industry leaders are taking an active role in establishing a foundational framework for the co-definition of Lumen Sovereign's use-case parameters, rigorous security protocols, and necessary governance standards. This collaborative effort ensures the delivery of a sovereign frontier capability architected from inception to satisfy the stringent assurance requirements of regulated industries.
Technical architecture overview
To deliver Lumen Sovereign by our target of late 2026, Cosine is leveraging its allocation of UK sovereign compute to upcycle a powerful open-weight model. Rather than pre-training from scratch, Cosine’s engineering focus is on rapid, high-impact capability scaling:
- Applying continual pre-training and expert expansion to the 256k-context sparse Mixture-of-Experts (MoE) architecture, targeting a long-horizon agentic coding model within a 500,000 H100 GPU-hour compute budget
- Enhancing the MoE architecture for highly specialised, domain-specific reasoning
- Implementing lightweight instruction recovery so the model maintains state-of-the-art proficiency in coding, tool use, and complex agentic workflows
Purpose-built for high-stakes workflows, Lumen Sovereign will be designed to tackle workflows where the combination of sensitivity and complexity has historically made AI adoption impossible.
The training regimen is strategically allocated, dedicating approximately one-third of the total compute time to optimising performance on complex software engineering problems, with the remaining two-thirds focused on broader, mission-critical use cases such as cybersecurity adversarial testing, KYC and AML alert investigations, and clinical trial coordination.
When complete, Lumen Sovereign will be capable of deployment entirely within a customer’s own infrastructure – including fully air-gapped environments – with absolutely no external data transfer required.
Making Britain an AI maker
By bringing together the backbone of UK critical infrastructure to co-design Lumen Sovereign, we are taking a massive step toward technological independence, ensuring that Britain remains an AI maker, not just an AI taker.