BrowseComp-ZH
Search · 2025-04-27
BrowseComp-ZH is a Chinese-web counterpart to BrowseComp: 289 multi-hop questions whose short, verifiable answers require browsing and reasoning over the Chinese web, written so that the answer is hard to find but easy to check. Scores are the share of questions answered correctly by a browsing agent; the agent framework and search tools are run-level settings. Higher is better.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-5.2 | 76.1 |
| Qwen3.5 397B A17B | 70.3 |
| Qwen3.5 122B A10B | 69.9 |
| Qwen3.5 35B A3B | 69.5 |
| Gemini 3 Pro Preview | 66.8 |
| DeepSeek-V3.2 | 65.0 |
| GPT-5 | 63.0 |
| Opus 4.5 | 62.4 |
| Kimi K2 Thinking | 62.3 |
| Qwen3.5-27B | 62.1 |
| Qwen3 Max Thinking (2026-01-23) | 60.9 |
| GPT-5 Mini | 49.5 |
| DeepSeek-V3.1 | 49.2 |
| MiniMax M2 | 48.5 |
| DeepSeek-V3.2-Exp | 47.9 |
| DeepSeek-V3.1-Terminus | 45.0 |
| gpt-oss-120b | 42.9 |
| Sonnet 4.5 | 42.4 |
| DeepSeek-R1-0528 | 35.7 |
| Kimi K2 Instruct 0905 | 22.2 |