Qwen open-sources BabelArena: same task took 25 tool calls in English, 201 in Telugu
Qwen's research compares one shopping task across two languages.
Qwen (@QwenDevs) posted on Monday that it is open-sourcing BabelArena research. In the test, Qwen-3.8-Max did the same shopping task in English and Telugu with the same result, but needed 25 tool calls in English versus 201 in Telugu. The post drew about 800 likes.
The post does not explain why the gap is so large or whether other languages or models were tested.
Why it mattersThe outcome matched, but the cost in tool calls did not, which hints that agents may be much less efficient outside English. Look at the released research for details.
