mGPT-13B: benchmark results
ai-forever's mGPT 13B, a multilingual GPT model pretrained on 600GB of MC4 and Wikipedia text across 61 languages from 25 families. Provider: Sber. Released 2023-04-18. Access: Open.
Aggregate Rating 1239 ± 21, rank #1523 of 1635 rated models, from 26 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MERA Original Paper - MathLogicQA | 26.3 | Accuracy (%; 5-shot, log-likelihood) | 72.2 |
| MERA Original Paper - SimpleAr | 2.3 | Exact Match (%; 5-shot, greedy generation) | 66.7 |
| MERA Original Paper - ruMultiAr | 1.9 | Exact Match (%; 5-shot, greedy generation) | 66.7 |
| Open Korean LLM - Ko-GPQA-Diamond | 24.24 | Normalized accuracy (%) | 62.4 |
| MERA Original Paper | 19.6 | Total Score (%; mean of 17 problem-solving and exam tasks, M | 44.4 |
| MERA Original Paper - ruOpenBookQA | 25 | Accuracy (%; 5-shot, log-likelihood) | 36.1 |
| Open Korean LLM - Ko-EQ-Bench | 0 | EQ-Bench score | 24.5 |
| MERA Original Paper - ruModAr | 0 | Exact Match (%; greedy generation) | 19.4 |
| Open Korean LLM - Ko-IFEval | 17.1 | Strict accuracy, prompt/instruction mean (%) | 12.4 |
| MERA Original Paper - ruMMLU | 23.5 | Accuracy (%; 5-shot, log-likelihood) | 11.1 |
| MERA Original Paper - ruWorldTree | 23.2 | Accuracy (%; 5-shot, log-likelihood) | 11.1 |
| Open Korean LLM - KorNAT-CKA | 15.18 | Normalized accuracy (%) | 8.4 |
Interactive version: theaggregate.ai/model?slug=mgpt-13b · How It Works · Data refreshed daily, snapshot 2026-11-05.