Celtic languages — including Cornish, Irish, Scottish Gaelic and Welsh — are the oldest living languages of the U.K. To empower their speakers, the UK-LLM sovereign AI initiative is building a new language model based on NVIDIA Nemotron that can reason in both English and Welsh, spoken today by around 850,000 people in Wales.
High-quality AI reasoning in Welsh will help deliver critical public services — from healthcare to education and legal resources — in the native language of communities.
“I want every corner of the U.K. to harness the benefits of artificial intelligence. By enabling AI to reason in Welsh, we’re making sure public services — from healthcare to education — are accessible to everyone, in the language they live by,” said U.K. Prime Minister Keir Starmer. “This is a powerful example of how advanced AI, trained on the U.K.’s most capable supercomputer in Bristol, can serve the public good, protect cultural heritage and unlock opportunity nationwide.”
Building on UK-LLM’s Mission
Launched in 2023 as BritLLM and led by University College London, the UK-LLM project has already released two language models for U.K. languages. Its latest Welsh model — developed in collaboration with Bangor University and NVIDIA — supports the Welsh government’s Cymraeg 2050 strategy to grow active use of the language and reach one million speakers by mid-century.
U.K.-based AI cloud provider Nscale will host the model through an API, giving developers and institutions an easy way to integrate Welsh-language reasoning into their services.
“The aim is to keep Welsh alive, evolving, and relevant in modern life,” said Gruffudd Prys, senior terminologist and head of the Language Technologies Unit at Bangor University’s Canolfan Bedwyr. “AI can support second-language learning as well as help native speakers refine their skills.”
Expanding Accessibility
The model opens new possibilities for Welsh-speaking communities by enabling public institutions and businesses to translate content and offer bilingual chatbot services. Healthcare providers, educators, broadcasters, retailers and restaurants could all ensure that written content is as accessible in Welsh as it is in English.
Beyond Welsh, the UK-LLM team plans to apply the same methodology to other U.K. languages such as Cornish, Irish, Scots and Scottish Gaelic, and to collaborate internationally on models for languages across Africa and Southeast Asia.
“This collaboration with NVIDIA and Bangor University enabled us to build training data and train a new model in record time,” said Pontus Stenetorp, professor of natural language processing and deputy director at UCL’s Centre for Artificial Intelligence. “Our goal is to use the lessons learned from Welsh to support other minority languages across the U.K. and worldwide.”
Harnessing Sovereign AI Infrastructure
The Welsh model is built on NVIDIA Nemotron, a family of open-source models with open weights, datasets and recipes. UK-LLM leveraged the 49B-parameter Llama Nemotron Super model and the 9B-parameter Nemotron Nano model, post-training them on Welsh-language data.
Because Welsh has far fewer digital resources than English or Spanish, the team generated a large dataset by using NVIDIA NIM microservices for gpt-oss-120b and DeepSeek-R1 to translate more than 30 million entries from Nemotron’s open datasets into Welsh.
Training and translation workloads were powered by NVIDIA DGX Cloud Lepton and hundreds of NVIDIA GH200 Grace Hopper Superchips on Isambard-AI, the U.K.’s most powerful supercomputer, backed by £225 million in government investment and hosted at the University of Bristol.
Capturing Nuances With Expert Evaluation
Bangor University, based in Gwynedd — the county with the highest proportion of Welsh speakers — provided cultural and linguistic expertise. Prys and his team helped verify translated datasets and assess how the model handled Welsh-specific challenges, such as initial consonant mutations depending on surrounding words.
The Welsh training data, evaluation sets and model itself are expected to be released for use by enterprises and the public sector, supporting research, model development and new applications.
“It’s one thing to have AI that works in Welsh — it’s another to make it open and accessible,” said Prys. “That subtle distinction is what determines whether people will use the technology.”
A Foundation for Multilingual AI
The approach behind the Welsh model can serve as a blueprint for multilingual AI development worldwide. Nemotron models, datasets and recipes are freely available, and packaged as NVIDIA NIM microservices to run cost-effectively across environments, from laptops to the cloud.
Enterprises across Europe will also be able to run sovereign, open models on Perplexity’s AI-powered search engine.
