T-pro-it-2.0: benchmark results
T-Bank's open-weight T-pro-it-2.0 (July 2025), a Russian-focused continual pretrain and alignment of Qwen3-32B, trained on 40B tokens of instruction data with about a third devoted to reasoning. Provider: T-Bank. Released 2025-07-17. Access: Open.
Aggregate Rating 1622 ± 26, rank #380 of 1656 rated models, from 34 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MERA - ruWorldTree | 99.24 | Accuracy (%) | 92.1 |
| MERA - CheGeKa | 51.79 | F1 (%) | 91.9 |
| MERA - PARus | 93.2 | Accuracy (%) | 89.3 |
| MERA - ruHHH | 87.64 | Accuracy (%) | 88.8 |
| MERA - ruCodeEval | 54.27 | pass@1 (%) | 87.6 |
| MERA - ruTiE | 87.6 | Accuracy (%) | 87.1 |
| MERA - ruOpenBookQA | 94 | Accuracy (%) | 86.7 |
| MERA - ruHateSpeech | 86.42 | Accuracy (%) | 85.7 |
| MERA - MaMuRAMu | 85.08 | Accuracy (%) | 83.8 |
| MERA - ruMMLU | 78.97 | Accuracy (%) | 82.9 |
| MERA - ruMultiAr | 48.73 | EM (%) | 81.4 |
| MERA | 0.66 | Total score | 80.8 |
Interactive version: theaggregate.ai/model?slug=t-pro-it-2-0 · How It Works · Data refreshed daily, snapshot 2026-10-20.