Qwen AI needs 8 times more steps in Telugu

Alibaba's Qwen-3.8-Max AI took 201 steps for a Telugu shopping task and 25 in English, the Qwen team wrote on the 26th of October 2026. 201 divided by 25 is about 8, so it needed about 8 times more steps.

Both runs found the same 4 products. The steps are tool calls, actions such as a product search that the AI agent makes.

Why Telugu took longer

In Telugu, a language of southern India, the AI ran 158 product searches against 5 in English. 126 of those searches came back empty, the paper says.

Across the 5 models, 53.7 to 68.7 percent of searches in less common languages came back empty. In English it was 6.1 to 17.1 percent.

What BabelArena tests

The example comes from BabelArena, a free test set from Alibaba researchers. It has 699 tasks in 13 areas of life and 23 languages.

The paper tested Claude Opus 4.8, GPT-5.6 Terra, Gemini 3.1 Pro, Qwen-3.7-Max and Qwen-3.8-Max. Qwen-3.8-Max scored best overall, though no model won every test.

Languages with less online text used 1.69 to 2.06 times as many input tokens as English, the paper found. A token is a small piece of a word.

Below we break down the numbers

In a Belarusian task, Claude Opus 4.8 needed 121 steps against 25 in English, by the paper's count. That is about 4.8 times more, by our arithmetic.

What is confirmed and what is not

The Telugu and Belarusian runs are single examples chosen by the authors. Averages across many tasks show smaller gaps.

Key facts

BabelArena has 699 tasks in 23 languages.

Alibaba researchers released it as free, open code.

One Telugu task took 201 steps against 25 in English.

When agents switched language, 91.2 percent of switches went to English.