MegaBeam-Mistral-7B-512k: benchmark results
Amazon's 7B long-context model, fine-tuned from Mistral-7B-Instruct-v0.2 to handle up to 524,288 tokens of context for retrieval and reasoning over long documents. Provider: Amazon. Released 2024-07-30. Access: Open.
Aggregate Rating 1381 ± 16, rank #1287 of 1656 rated models, from 34 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| DropKV Reference - RULER 128K (full KV cache) | 89.02 | Accuracy (%) | 100 |
| Open Korean LLM - Ko-IFEval | 39.58 | Strict accuracy, prompt/instruction mean (%) | 81.2 |
| Open LLM Leaderboard - IFEval | 59.73 | Score | 71.6 |
| Open Korean LLM - Ko-EQ-Bench | 22.24 | EQ-Bench score | 67.6 |
| Open Korean LLM - KorNAT-Harmless | 63.46 | Normalized accuracy (%) | 63.1 |
| Open Korean LLM Leaderboard | 37.41 | Average Score (%) | 61.9 |
| Open Ko-LLM Leaderboard | 37.41 | Average (%) | 61.7 |
| Open Korean LLM - KorNAT-Helpful | 47.83 | Normalized accuracy (%) | 61.4 |
| Open Korean LLM - Ko-GPQA-Diamond | 23.74 | Normalized accuracy (%) | 55.6 |
| Open Portuguese LLM - ASSIN2 RTE | 89.37 | Macro F1 (%) | 49.9 |
| Open Korean LLM - Ko-GSM8K (flexible extract) | 17.97 | Exact match, flexible extract (%) | 49.1 |
| Open LLM Leaderboard - MuSR | 39.94 | Score | 41.4 |
Interactive version: theaggregate.ai/model?slug=megabeam-mistral-7b-512k · How It Works · Data refreshed daily, snapshot 2026-10-20.