Atlas

Benchmarks

← All benchmarks

OmnilingualGAIA2

Agents · 2026-08-09

OmnilingualGAIA2 extends four GAIA2 agent capabilities (search, execution, ambiguity and adaptability) to ten languages: Spanish, French, Portuguese, Indonesian, Italian, Turkish, Chinese, German, Hindi and Japanese. Tasks and the simulated universe are both in the target language, and a verifier with a multilingual LLM judge grades actions and responses. The headline score macro-averages pass@1 over the four capabilities and then over the ten languages.

Top models (higher is better)

ModelScore
Muse Glimmer 30B60.7
Qwen3.6 27B45.0
Gemma 4 31B IT37.5
Loading Atlas data…