MERA - BPS: leaderboard
MERA is the Russian open benchmark run by the AI Alliance Russia; this is its BPS task: Balanced Parentheses Sequence: decide whether a bracket sequence is balanced, a pure formal-structure probe with no world knowledge in it. Each MERA task keeps its own native metric, so the boards are not on one shared scale.
Metric: Accuracy (%). Source: mera.a-ai.ru. Saturation forecast: Estimated already saturated. 631 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 72B Instruct | 100 | #454 |
| 2 | Qwen 3.5 27B | 100 | #187 |
| 3 | Qwen 3.5 122B A10B | 100 | #173 |
| 4 | Claude Opus 4.5 | 100 | #83 |
| 5 | Qwen 2.5 32B Instruct | 100 | #498 |
| 6 | Median Human | 100 | |
| 7 | GPT-5.4 | 99.9 | #79 |
| 8 | DeepSeek V3 | 99.9 | #321 |
| 9 | Qwen 3.6 27B | 99.9 | #170 |
| 10 | Gemini 2.0 Flash | 99.8 | #342 |
| 11 | GLM-5.1 | 99.8 | #97 |
| 12 | Gemma 4 26B A4B (IT) | 99.8 | #297 |
| 13 | Gemini 2.0 Flash Lite | 99.8 | #450 |
| 14 | Qwen 2.5 Max | 99.8 | #416 |
| 15 | Gemma 4 31B (IT) | 99.7 | #176 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mera-bps · How It Works · Data refreshed daily, snapshot 2026-10-20.