DeepSeek V4-Flash-0731 open weights tested
An open model with 13 billion active parameters, with documentation clearly stating that Welsh is not among its supported languages.
The facts
| Developer | DeepSeek |
|---|---|
| Released | 31 July 2026 (second post-trained version of April’s V4-Flash preview, using the same architecture) |
| Architecture | MoE with 284 billion parameters in total / 13 billion active; a one-million-token context window |
| Licence | MIT, with weights available without access restrictions |
| API pricing | $0.14 / $0.28 per million tokens — around ten times cheaper than the main closed models |
Welsh: the documented gap
DeepSeek’s documentation explicitly describes its language support. According to its language support guide, English and Chinese receive the strongest support, with some support for major European and East Asian languages. ‘Welsh and other low-resource Celtic languages’ do not appear in the documented support tiers. This provides useful context for interpreting Welsh tests.
Why include it in the research?
The combination of 13 billion active parameters, an MIT licence and open weights makes it practical to study and fine-tune the model locally on Welsh. The resources needed to run it are within reach of some independent projects, while the large total parameter count retains broad knowledge. It could therefore offer a way to test an approach similar to Jupiter-N outside the Nemotron family. It is on the model list for our experiments.
CymraegBench v0.1 result
Overall score 61.63 · Welsh reasoning 63.14 · practical Welsh 60.11 · translation BLEU 64.95. DeepSeek placed fifteenth, but its translation score was higher than its overall score. The later DeepSeek V4.1 Flash scores 7.82 points higher overall. See the full leaderboard.