Masked diffusion
language models.
A family of sovereign, UK-built models that generate text in parallel — faster and cheaper to run than the models you rent today.
Coming soonThree sizes, named for British rivers.
One family, smallest to largest — so you run the right model for the job, not the biggest one for every job.
Wye
The small, fast model. Efficient enough to run high-volume document work at low cost — built for scale, not headlines.
Severn
The balanced workhorse. Everyday capability for regulated metadata, classification and reasoning across the document estate.
Thames
The largest model. Frontier-scale reasoning for the hardest regulated problems — sovereign, and trained on UK-governed data.
Why masked diffusion.
Parallel by design
Masked diffusion models generate by unmasking text across the whole sequence at once — not one token at a time.
Cheaper to run
Parallel decoding cuts the per-token cost a buyer actually pays — where a sovereign provider’s economics are won.
Capital-efficient
Built by adapting open models rather than training from scratch — and on regulated data, inside the UK.
Want to run a sovereign model on your data?
The models are in development. If you're a regulated-industry team who wants early access — or to shape one on your domain — talk to us.