31 July 2026

Three new open models, but no Welsh data

Three large open-weight models were released in seven weeks. Welsh does not appear in their public documentation.

It has been a busy summer for open-weight models. With DeepSeek V4-Flash-0731 released today, this is the third major model from China since mid-June:

The three models have similar aims: a one-million-token context window, agentic ability and strong coding skills. Their main languages are Chinese and English. All three models’ weights can be downloaded — 1.56TB for K3 — and two have an MIT licence.

What about Welsh?

After searching documentation, model cards and analyses of all three, we found no claim, benchmark or reference to Welsh or the Celtic languages. A guide to DeepSeek’s supported languages also notes that Welsh and other lower-resource Celtic languages do not appear in the documented support tiers.

That does not mean they cannot generate Welsh. Large web corpora contain some Welsh material, and our evaluations section will test that ability. But there is no evidence that the labs themselves have measured it, so capability could deteriorate between versions unnoticed.

Why is this still an opportunity?

For Welsh work, the most important figure in the table is DeepSeek’s 13 billion active parameters under an MIT licence. Although the model has 284 billion parameters in total, only 13 billion are active at a time. As a result, it can run on a workstation with 128GB of memory — the same type of machine used in our fine-tuning experiments. The MIT licence permits fine-tuning, publication and redistribution without additional permission.

In other words, the weights and licences make a community project to fine-tune and assess a model on Welsh more practical. This gives researchers more options for measuring Welsh directly.

Sources: the three labs’ HuggingFace and GitHub announcements; analyses by MLQ, DataNorth and MarkTechPost; deepseekai.guide language guide. Full details in the profiles of the evaluations section.