Meta's largest open-weight model, rivaling GPT-4 class performance with full open weights for commercial and research use.
| Benchmark | Description | Score | Category | Bar |
|---|---|---|---|---|
| MMLU | Massive Multitask Language Understanding | 88.6% | knowledge | |
| HumanEval | Code generation accuracy | 89% | coding | |
| MATH | Mathematical reasoning | 90% | math | |
| GSM8K | Grade school math | 93% | math | |
| GPQA | Graduate-level science Q&A | 51.2% | reasoning | |
| HellaSwag | Common sense reasoning | 86.9% | reasoning | |
| SWE-bench | Software engineering problems | 38.3% | coding | |
| ARC-C | AI2 Reasoning Challenge | 95% | reasoning |