Mixtral 8x22B: benchmark results
Mistral's Apache-2.0 sparse-MoE base model (141B total/39B active, 64K context) released April 2024; this row is the pretrained, non-instruct weights. Provider: Mistral. Released 2024-04-17. Access: Open.
Aggregate Rating 1506 ± 13, rank #766 of 1656 rated models, from 184 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| HELM NaturalQuestions (Closed) | 47.76 | F1 (%) | 95.6 |
| HELM MMLU - Virology | 59.64 | EM | 95.5 |
| HELM MMLU - Security Studies | 86.53 | EM | 93.6 |
| HELM MMLU - World Religions | 90.06 | EM | 91 |
| CyberSecEval-3 Insecure Autocomplete | 25 | Insecure code generation (%) | 90 |
| SEAL - Fortress | 56.06 | Score | 88.6 |
| SynthPAI | 72 | Average accuracy in % | 88.2 |
| HELM NarrativeQA | 77.87 | F1 (%) | 87.2 |
| HELM Lite - LegalBench - Function of Decision Section | 44.14 | Exact Match (%) | 85 |
| HELM Lite - WMT 2014 (de-en) | 20.86 | BLEU-4 (%) | 85 |
| HELM Lite - LegalBench | 70.79 | Exact Match (%) | 84.4 |
| HELM Lite - LegalBench - Abercrombie | 76.84 | Exact Match (%) | 83.3 |
Interactive version: theaggregate.ai/model?slug=mixtral-8x22b · How It Works · Data refreshed daily, snapshot 2026-10-20.