Hermes 3 is a fine-tuned conversationalist built on Meta's Llama 3.1 8B base, shaped by NousResearch to follow instructions precisely and engage in structured, multi-turn dialogue. It handles roleplay, reasoning chains, and agentic tasks with more consistency than its base model, leaning into explicit system prompt adherence. At 8B parameters, it punches reasonably well for its size but will show limits on complex multi-step reasoning compared to larger siblings.
| Benchmark | Score | Type | Recorded |
|---|---|---|---|
| IFEval | 61.7 | accuracy | 6d ago |