41 more languages were tested on the 20 everyday exchanges, with Gemma 4
E2B and E4B. Two independent model readers read every answer, and Gemma 4
31B a third time wherever their verdicts could change the outcome (it
agreed with the hand reading of English, Spanish, German and Hindi on 75
of 80 verdicts). A language is ready on a model at 15 of 20 usable
sentences and 60 of 80 good replies, and is offered with the smallest
model it is ready on.
- Ready with E2B: Arabic, Danish, French, Indonesian, Korean, Malay,
Persian, Portuguese, Russian, Swedish, Turkish, Urdu, Vietnamese (with
English, Spanish, German and Hindi: 17).
- Ready with E4B: Chinese, Dutch, Gujarati, Hebrew, Italian, Japanese,
Marathi, Norwegian, Polish, Punjabi, Tagalog, Telugu, Thai, Ukrainian.
- Not yet, each with what went wrong: Amharic, Bengali, Czech, Finnish,
Greek, Hausa, Hungarian, Kannada, Malayalam, Romanian, Swahili, Tamil,
Yoruba, Zulu.
Every answer and verdict is in eval/results-2026-10-01-languages.json. No
native speaker has read these yet. The committed catalogue's test now
checks every ready language and the larger model's cover.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
f71fa28447