The UK-LLM initiative is developing an AI model capable of reasoning in English and Welsh, with the aim of supporting the delivery of public services such as healthcare, education and legal resources in Welsh, which is spoken by approximately 850,000 people in Wales.
The new model is based on the open-source NVIDIA Nemotron family and was developed by University College London in collaboration with Bangor University and NVIDIA. The model, along with Welsh training and evaluation datasets, is expected to be made available for use by institutions, businesses and the public sector, enabling further research, the training of new models and the development of applications.
Supporting Public Services and Local Languages
UK Prime Minister Keir Starmer said that enabling AI to reason in Welsh could make public services, from healthcare to education, available to everyone in the language in which they live. The project is linked to the Welsh Government’s efforts to promote active use of the language under the Cymraeg 2050 initiative, which aims to reach one million Welsh speakers by 2050.
The model could help public institutions and businesses operating in Wales translate content or provide bilingual chatbots. This includes healthcare providers, teachers, media outlets, retailers and restaurant owners, helping them provide written content in Welsh alongside English.
The initiative was established in 2023 under the name BritLLM and is led by University College London. It previously released two models for languages used in the United Kingdom. The UK-LLM team intends to apply the same methodology to other languages, including Cornish, Irish, Scottish and Scottish Gaelic, and also plans to work with international organizations to develop models for languages in Africa and Southeast Asia.
Training Data and Computing Infrastructure
The development team used the Llama Nemotron Super model, which has 49 billion parameters, and the Nemotron Nano model, which has 9 billion parameters, and then carried out further training on them using Welsh-language data. Because the amount of available Welsh data is limited compared with English and Spanish, the team used NVIDIA NIM microservices for gpt-oss-120b and DeepSeek-R1 to translate Nemotron’s open datasets, which contain more than 30 million entries, from English into Welsh.
The translation and training work was carried out using a GPU cluster through the NVIDIA DGX Cloud Lepton platform, in addition to hundreds of NVIDIA GH200 Grace Hopper Superchips on the Isambard-AI supercomputer. The computer is located at the University of Bristol and is supported by a £225 million investment from the UK government.
Linguistic Review and Developer Access
Bangor University supports the project with its linguistic and cultural expertise, while Gruffudd Prys and his team review the accuracy of the machine-translated training data and manually translated evaluation data. The team also assesses the model’s ability to handle subtle features of Welsh, such as changes to consonants at the beginnings of words depending on neighboring words.
Nscale, a UK-based cloud AI provider, will make the new model available to developers through an application programming interface. The UK-LLM team believes that the framework used could form the basis for developing multilingual models, making use of Nemotron models and their publicly available data and recipes.