bigstral-12B-32k: benchmark results
Abacus.AI's bigstral-12b-32k, a 12B self-merge of Mistral-7B-Instruct-v0.2 made by interleaving its layers via mergekit, built to offer 32K context where SOLAR only had 4K. Provider: Abacus.AI. Released 2024-03-06. Access: Open.
Aggregate Rating 1377 ± 17, rank #1306 of 1656 rated models, from 29 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Korean LLM - KorNAT-Helpful | 50.95 | Normalized accuracy (%) | 88.1 |
| Open LLM Leaderboard - MuSR | 45.6 | Score | 86.1 |
| Open Korean LLM - KorNAT-Harmless | 66.02 | Normalized accuracy (%) | 83.8 |
| Open Chinese LLM - TruthfulQA MC | 58.81 | Accuracy (%) | 78.7 |
| Open Korean LLM - Ko-IFEval | 36.57 | Strict accuracy, prompt/instruction mean (%) | 75.8 |
| Open Korean LLM - Ko-EQ-Bench | 28.25 | EQ-Bench score | 75.7 |
| Open Portuguese LLM - ASSIN2 RTE | 90.72 | Macro F1 (%) | 60.1 |
| Open Portuguese LLM - BLUEX | 51.74 | Accuracy (%) | 50.4 |
| Open Korean LLM Leaderboard | 35.18 | Average Score (%) | 50.2 |
| Open Portuguese LLM - ENEM | 61.65 | Accuracy (%) | 50.2 |
| Open Ko-LLM Leaderboard | 35.18 | Average (%) | 50 |
| Open Portuguese LLM - OAB Exams | 41.78 | Accuracy (%) | 49.2 |
Interactive version: theaggregate.ai/model?slug=bigstral-12b-32k · How It Works · Data refreshed daily, snapshot 2026-10-20.