University College London, NVIDIA and Bangor University have developed a UK-LLM model trained on the Isambard-AI supercomputer to bring artificial intelligence reasoning to Welsh and other U.K. languages. The project is part of a sovereign Artificial Intelligence initiative and aims to make public services more accessible in Welsh by supporting healthcare, education and legal resources. U.K. prime minister Keir Starmer described the effort as a way to protect cultural heritage and extend benefits of artificial intelligence across the country, and the work aligns with the Welsh government's Cymraeg 2050 goal of one million speakers by 2050.
The new Welsh model is based on NVIDIA Nemotron open-source technology, using the 49-billion-parameter Llama Nemotron Super model and a 9-billion-parameter Nemotron Nano model that were post-trained on Welsh-language data. To overcome limited Welsh training material, the team used NVIDIA NIM microservices alongside gpt-oss-120b and DeepSeek-R1 to translate NVIDIA Nemotron open datasets, producing a training set with over 30 million English-to-Welsh entries. Training and translation workloads ran on NVIDIA DGX Cloud Lepton and hundreds of GH200 Grace Hopper Superchips on Isambard-AI at the University of Bristol, leveraging significant government-backed infrastructure to accelerate development.
Bangor University and Canolfan Bedwyr contributed linguistic and cultural expertise, verifying machine-translated data and assessing the model's handling of Welsh-specific features such as initial-consonant mutations. Nscale will provide access to the model via an API for developers, and the team expects to make the model and Welsh datasets available for enterprise and public sector use to support further research and application development. The UK-LLM effort intends to apply the same methodology to other minority U.K. languages-including Cornish, Irish, Scots and Scottish Gaelic-and to work with international partners to build models for languages in Africa and Southeast Asia.
