howat.
Rearchitecting LLMs for a more capable,
more efficient kind of intelligence.
We’re a UK AI research lab developing more compute-efficient language model architectures.
Results so far
- Original model · GPU joules per correct benchmark answer
- 122.7J
- Re-architected · GPU joules per correct benchmark answer
- 109.4J
- GPU energy per correct benchmark answer
- −10.8%
From research to deployment.
In our first funded year, we aim to develop and retrain a 32–70B mixture-of-experts model with our architecture integrated internally, and launch UK-hosted inference for enterprise and public sector use.
Our ambition is more useful AI with less energy, supporting the UK’s net-zero targets and strengthening domestic AI capability.
Research overview
We measure answer quality against energy and computation, testing architectural changes across model sizes.
The funded programme will compare original and re-architected models on held-out language, reasoning, coding and multi-turn tasks, with independent evaluation and pilots for enterprise and public sector use.
Our year-one target is at least 10% lower whole-service energy per successful task at matched answer quality. We will measure inference costs and associated emissions, including the additional retraining energy.