Higgs-Llama-3-70B: benchmark results
Boson AI's role-play-focused tune of Llama 3 70B, combining SFT with iterative preference optimization for character and system-prompt adherence. Provider: Boson AI. Released 2024-06-05. Access: Open.
Unified ELO 1609 ± 1, rank #238 of 1392 rated models, from 14 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Open Chinese LLM - CMMLU | 75.33 | Accuracy (%) | 99.1 |
| Open Chinese LLM - GSM8K | 71.65 | Accuracy (%) | 98.8 |
| Open Chinese LLM Leaderboard | 70 | Average Score (%) | 96.7 |
| Open LLM Leaderboard - GPQA | 15.55 | Score | 91.9 |
| Open Chinese LLM - ARC Challenge | 61.26 | Accuracy (%) | 90.5 |
| Open Chinese LLM - C-Eval Semantic | 88.16 | Accuracy (%) | 87.4 |
| Open LLM Leaderboard - MMLU-Pro | 43.35 | Score | 87.1 |
| Open Chinese LLM - HellaSwag | 68.31 | Accuracy (%) | 86.9 |
| Open LLM Leaderboard - BBH | 45.9 | Score | 86 |
| Open LLM Leaderboard - MuSR | 15.52 | Score | 83.2 |
| Open Chinese LLM - WinoGrande | 67.4 | Accuracy (%) | 82.5 |
| Open LLM Leaderboard - MATH Level 5 | 25.23 | Score | 78.2 |
Interactive version: theaggregate.ai/model?slug=higgs-llama-3-70b · How It Works · Data refreshed daily, snapshot 2026-09-05.