News
News about artificial intelligence and Welsh—models, research, policy and projects moving the field forward.
From dictionary to speech: 3 new lessons for Welsh AI
New work on Irish, Basque and rare-language archives offers three concrete experiments for Welsh AI speech, data and testing.
More than speaking Welsh: data, law and real testing
A new programme for bilingual law and work on lower-resource languages show that AI readiness means authoritative data, governance and real testing.
Behind the score: five new models, five different profiles
Five new CymraegBench runs show that a high overall score alone does not guarantee consistency, instruction following or neutrality.
Welsh punctuation on your device: an open model, a focused test
Two new open models restore Welsh punctuation locally; new research shows how to test tokenisation, language switching and on-device use.
From Kolibri to Wales: how much would a sovereign Welsh AI model cost?
A Welsh model the size of Kolibri would cost around £2.5m–£3.5m in computing; a practical sovereign programme, including data, people, testing and deployment, would be closer to £10m.
From principle to testing: building Welsh into AI
A UNESCO toolkit and a new test of models’ grammatical ability show why Welsh needs to be built in from the start, and every checkpoint tested.
Four new models, one lesson: the details matter more than the ranking
Opus 5.5 ranks second on CymraegBench, but four new evaluations show why the profile behind the score matters more than the ranking.
From meetings to evidence: a clearer map for Welsh AI
There was no new Welsh model, but practical use in the Senedd and work on lower-resource languages show how to move from data to a trustworthy test.
BuzzASR: a new Welsh speech model to test
A specialist Welsh speech recognition model is available to download, with a large reduction in errors on two test sets—but testing on real-world speech is still to come.
Four grants to build Welsh AI’s foundations
Wales is investing in data, maps, open knowledge and language infrastructure — four complementary areas for the next generation of Welsh technology.
High correction scores, but explanation is another matter
Three new models achieve strong Welsh scores, but an audit of 3,712 answers reveals a large gap between correction and explanation.
Qwen3.8 27B: what lies behind the score?
Qwen3.8 27B scores 70.47 on CymraegBench, but the gap between Welsh reasoning and practical Welsh shows why results need close examination.
Local language and living voices: new lessons for Welsh AI
BavGround and Easper show why local knowledge must be tested and recordings selected for linguistic value, not just audio clarity.
Luth-2: a small model showing a way forward for Welsh
Luth-2 shows how post-training a small model for one language can deliver large gains—and why tests written in Welsh, reflecting local life and including human judgement, are needed.
When the judges called a learner AI
Romy Wood denies using AI for Y Dderwen, the judges stand by their view, and a new Yiddish model shows why data provenance and native-language tests are essential.
Three new open models, but no Welsh data
Kimi K3, GLM-5.2 and DeepSeek V4-Flash-0731—three open-weight models in seven weeks, but no Welsh results. Open licences could create new opportunities for local fine-tuning.
Welsh Government AI summary: bilingual transcription and more Welsh data
Bilingual Teams transcription, Welsh Copilot, a data partnership with OpenAI, and grants for Iriaith, Mapio Cymru, Wici-Fflics and CEFR: the Government's 2025–26 summary, alongside the Welsh Language Commissioner's four AI principles.
Correction and explanation: findings from the Coleg Cymraeg report
A study by UWTSD and Coleg Cymraeg Cenedlaethol finds that ChatGPT and Copilot successfully correct many Welsh errors, but give unreliable explanations, use unnatural idiom and generate fabricated references. The practical finding: one sentence in the prompt can prioritise Welsh sources.
GPT-5.6 Sol: a new model, but no measure of its Welsh
The GPT-5.6 family (Sol, Terra and Luna) has been released, but the system cards' multilingual tables contain no Welsh data.
Jupiter-N-120B: a large model trained on Welsh, available to everyone
Locai Labs and UCL have post-trained Nemotron-3-Super-120B on Welsh, British cultural context and agentic tasks. The datasets are openly available, alongside rare Welsh benchmark results and a ‘Forget-Me-Not’ framework to preserve existing capabilities.