Grok 4.20 commercial tested
xAI says Grok works in dozens of languages, but does not name Welsh or publish results for it.
The facts
| Developer | xAI |
|---|---|
| Current version | Grok 4.20 — system card dated 7 April 2026 |
| Licence | Proprietary (API) |
| Version note | Names are inconsistent across sources. Some external sources describe a ‘4.3’ model from the same period, and there are unconfirmed reports of 4.6 (1.5T parameters) and 4.7 (2.1T) for August 2026. We follow the official system card. |
Welsh: not named
Grok 4.20’s system card names five languages — English, Chinese, Arabic, Hindi and Spanish — ‘plus several dozen others with varying capability’. Welsh is not named, and no Welsh benchmark has been published. Grok Voice says it supports over 25 languages, but there is no confirmation that Welsh is among them.
This leaves Welsh in an unspecified category: the model may have some ability, but nobody has stated how much. The purpose of the leaderboard is to turn that ‘varying capability’ into a measurable result.
CymraegBench v0.1 result
Overall score 58.13 · Welsh reasoning 62.38 · practical Welsh 53.89 · translation BLEU 64.20. Grok placed sixteenth, without refusing a request. See the full leaderboard.