BEAR-Bench: leaderboard
Bilingual reasoning over text-dense professional documents in Russian and English, spanning business and science. Distinct from document-extraction benchmarks: the answer is not in a field to be pulled out, it has to be reasoned to.
Metric: Overall (%). Source: arxiv.org. Saturation forecast: Around January 2027. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 397B A17B | 75.4 | #146 |
| 2 | Gemini 3.1 Pro (Preview) | 75.1 | #58 |
| 3 | Gemini 2.5 Pro | 67.4 | #152 |
| 4 | Qwen 3.6 Plus | 65 | #120 |
| 5 | Claude Sonnet 4.6 | 62.7 | #90 |
| 6 | Claude Opus 4.6 | 61.5 | #63 |
| 7 | Qwen 3.5 27B | 60.8 | #187 |
| 8 | Qwen 3.5 9B | 49.6 | #379 |
| 9 | Gemini 2.5 Flash | 49 | #245 |
| 10 | Qwen 3.5 4B | 48.6 | #475 |
| 11 | Qwen 3 VL 4B (Thinking) | 34.5 | #480 (Qwen 3 VL 4B) |
| 12 | Qwen 3 VL 8B Instruct | 25.4 | #420 |
| 13 | Gemma 4 31B (IT) | 23.5 | #176 |
| 14 | Qwen 3.5 0.8B | 10.6 | #1424 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=bear-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.