Cosine

Building the UK’s first fully sovereign frontier AI model is no small feat, but the roadmap for Lumen Sovereign is locked in. We’re pushing the boundaries of architecture at Cosine to deliver a world-class LLM, and the next few months are critical. – Baseline & architecture validation (June-Aug 2026): We are currently testing our Mixture-of-Experts scaling pilots. Our lead configuration targets 192-256 experts with top-6 routing. We are promoting configurations based on rigorous loss analysis, SWE evals, long-context checks, and expert routing stability. – Model enlargement (Aug-Sept 2026): This is where the heavy lifting happens. We'll run the main expert-expansion continued pretraining phase to produce our 1 trillion-parameter intermediate checkpoint (based on the 256K source architecture) with 49.2B active parameters. – Agent post-training (Sept-Oct 2026): We'll run supervised fine-tuning on high-quality agentic coding demonstrations, followed by asynchronous reinforcement learning on long-horizon executable coding tasks. Building frontier AI requires pushing the absolute boundaries of model architecture. By Q4, our coalition partners will gain early access to these checkpoints, paving the way for a Q1 2027 commercial launch, available in the cloud, VPC, or fully air-gapped on your own GPUs. I am deeply proud of the relentless engineering going into this. It’s a big job, but we have the highly capable talent to do it. Keep an eye out over the coming weeks for an inside look into our lab.

We are incredibly proud to have Vodafone Business supporting our landmark industry coalition. Together, we are co-designing Lumen Sovereign to ensure the UK has a secure, self-reliant AI stack built for our critical national infrastructure and highly regulated enterprise environments.

An AI model has now found a zero-day, escaped its sandbox, moved laterally through internal systems and reached another company’s production environment. Any board still calling autonomous cyber offence a “future risk” is choosing denial. OpenAI’s incident should end the comforting fiction that attackers will need large teams of exceptional humans to conduct sophisticated operations. A capable model can work continuously, test thousands of paths, remember every failed attempt and combine weaknesses that no individual analyst would necessarily connect. There is no credible reason to believe hostile states, organised crime groups and well-funded adversaries will ignore this capability. They will use it, improve it and point it at banks, defence supply chains, telecoms networks and critical infrastructure. The defensive answer cannot be another dashboard, another annual penetration test or blind dependence on a hosted foreign model that may refuse a forensic request when an incident is already under way. Organisations will need the ability to attack themselves with comparable intelligence, inside their own perimeter, under their own control, before somebody else does it for real. That is why we built Cosine Red Teaming. In controlled evaluations, it has chained multiple exploits, identified zero-day vulnerabilities and, when technical routes narrowed, started moving towards social-engineering paths. We are deliberately rolling it out slowly because this is not a product that should sit behind a public signup form. The capability is too powerful, and the organisations using it need the governance, maturity and operational discipline to handle it safely. The next cyber divide will be brutally simple: organisations that discover their weaknesses at machine speed, and organisations whose adversaries discover them first. We are selecting a small number of trusted organisations for the next Cosine Red Teaming cohort. Those responsible for protecting defence, financial services, telecommunications, government or critical national infrastructure should speak to me directly.

Last week, our ML team stepped out of the office for a field trip to the Bristol Centre for Supercomputing (BriCS). They visited Isambard-AI to get a first-hand look at the infrastructure powering Lumen Sovereign, the first frontier AI model built entirely on UK soil. We’re proud to have the backing of the Sovereign AI fund, which has provided access to Isambard-AI, enabling us to take on this ambitious project and help keep UK AI at the cutting edge. A huge thanks to the BriCS team for the tour and their continued support. Now, back to training!

Are autonomous AI agents burning through your cloud API budgets? As engineering teams adopt agentic workflows, cloud token costs are scaling faster than forecast. For many enterprise teams, bringing AI infrastructure on-premise can become a more cost-effective path forward. We have modelled an illustrative scenario comparing the infrastructure economics of cloud-based coding agents with air-gapped sovereign GPUs operating in the UK. Key findings include: → Sovereign GPU deployments generally become more cost-effective from approximately 200 developers. → For a 2,000-person engineering team, a fully loaded on-prem deployment drops costs to ~£118–£174 per developer/month, significantly undercutting the estimated ~£230/month cloud average. → Owning your hardware can make compute expenditure more predictable and reduce exposure to usage-driven API costs. Read our full cost breakdown and deployment comparison: https://lnkd.in/eDpufA5P

In this conversation with Yang Li, co-founder and COO of Cosine, we discuss what sovereign AI actually means in practice. While the term has become increasingly common, Li argues that it comes down to three things: control, certainty, and resilience. Rather than relying on cloud-based AI services, Cosine builds foundation models that can run entirely inside a customer’s own infrastructure, including fully air-gapped environments where systems have no connection to the public internet. “Cosine is a model lab. We pre-train and post-train our own foundational models, and we produce the product layer as well, similar to Claude Code and OpenAI Codex. But fundamentally, we can deploy it in a fully air-gapped environment,” said Li. This ability has made Cosine particularly attractive to defence organisations and other highly regulated industries where sending sensitive information to a cloud provider is not an option. Customers include companies working with classified information as well as financial institutions that want greater control over how their data is handled. One obvious question is how an air-gapped AI system stays current. Li said Cosine doesn’t push continuous updates in the way cloud providers do. Instead, every customer agrees on an update schedule that matches the rest of its software infrastructure. Some organisations prefer weekly updates, while others only update periodically. Cosine also provides customers with the same tools its own engineers use, allowing them to fine tune and post-train models using their own internal data without exposing it outside their organisation.

Last week, the Cosine team was in Paris for the RAISE Summit. Sovereign AI was the number-one topic of conversation at RAISE, specifically, the need to hedge against single-provider or single-country dependencies to avoid a critical infrastructure kill switch. RAISE was unique. It was the first show we’ve attended that featured the entire stack in one room. European clouds, chip designers, model labs, and data centres – all pushing for a stronger, secure, and less dependent Europe. But there’s a harsh reality the European market has to address: the EU wants a seat at the table, but actual builders remain rare. Truly closing this gap demands sustained, on-the-ground investment. For Cosine, it was a fantastic week. We made incredible progress advancing relationships with some of Europe's most critical infrastructure providers. We are here to ensure that sovereign AI doesn't just become a buzzword, but a secure, built reality.