Dear SoTA,
Back in April, the UK’s Sovereign AI Fund announced the first group of British AI companies they were supporting. Cosine was one of them. We went to the fund with a clear and ambitious goal: to build a true frontier AI model from the ground up on UK soil. Recognising the stakes, they backed our vision with 500,000 GPU compute hours on Isambard-AI at the Bristol Centre for Supercomputing.
Access to one of the world’s most powerful supercomputers kickstarted things for us, but having the compute is only half the story. The more important question – one we discuss daily with leaders across UK industries – is why it actually matters for Britain to have its own models.
People often talk about “Sovereign AI” as just being about where the GPUs are. To us, it’s about control. It means institutions being able to run, look inside, and truly rely on these systems within their own secure environments, without worrying that a company elsewhere might suddenly change the rules or cut off access.
At Cosine, our focus is on building highly capable generalist LLMs and software engineering agents designed for these exact environments. A lot of our national infrastructure – such as defence, financial services, healthcare, and CNI – runs on code that doesn’t look much like the modern stacks dominating public training data. These systems depend on languages like COBOL, Fortran, and Ada. They also have highly specific use-case requirements.
Public models may excel on more fashionable frameworks, though they often struggle with the older, critical code these sectors actually use. Plus, for a lot of these organisations, sending data to a foreign API just isn’t an option. For them to really own the technology, the model has to run where they are, whether that is a managed cloud, a dedicated tenant, or a fully air-gapped environment.
That’s why we’re building Lumen Sovereign. We’re working with a group of UK partners in defence, finance, and tech to make a world-class model in the UK that’s designed for sensitive work and shaped by the people who will actually use it.
To date, we’ve made strong progress. In our latest pre-training run, we trained a 200B MoE model, which we have now finished instruct-tuning. We have also been experimenting with streamlined chat templates, identifying significant scope for simplification compared to existing solutions. Everything is proceeding as planned, which is highly encouraging.
Next, we will perform RL post-training on our 200B model, conducting thorough ablations and evaluations to identify the key drivers of performance. Following this, we will advance to the 1T+ model development phase.
Our partners will start testing these versions in the next couple of months, and we’re aiming for a full launch in early 2027. We’re making sure that from the start, organisations can run this on their own hardware or in their own private clouds.
The UK can either just keep buying AI from abroad, or we can build the secure systems our own institutions need. If you’re working on things like secure deployment, legacy systems, or cybersecurity, I’d love to chat.
Yours,
Alistair Pullen
Author biography:
Alistair Pullen is CEO & Co-Founder at Cosine. He leads the ML team, driving the company’s research direction, synthetic data curation, and model fine-tuning.
Write to the Society for Technological Advancement on letters@ilikethefuture.com.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.