Nous Hermes: The Fine-Tuned LLM Built for Instruction Following

Nous Hermes is a series of open-source LLMs produced by Nous Research, a small AI lab focused on fine-tuning open-source base models to improve instruction following, reasoning, and practical task performance. The Hermes series has built a strong following in the local LLM community for its balanced combination of capability and compliance.

What makes Hermes different

Hermes models are fine-tuned versions of base models like Llama and Mistral, trained on high-quality synthetic datasets produced by Nous Research. The key characteristics:

• Strong instruction following — better at following complex, multi-part instructions than many base fine-tunes
• Low refusal rate — less likely to refuse benign requests than heavily safety-tuned models like Claude
• Long context handling — maintains coherence across long conversations better than many alternatives
• JSON and structured output — reliable structured output without complex prompting
• Roleplay and persona — maintains consistent personas in extended interactions

The Hermes model lineup

Running Hermes with Ollama

ollama pull nous-hermes2         # Mixtral-based
ollama pull hermes3              # Llama 3.1 8B
ollama run hermes3

Hermes vs vanilla Llama

If you're already running Llama locally, Hermes offers a meaningful quality improvement for instruction-heavy tasks without changing your infrastructure. The fine-tuning adds instruction quality that the base Llama model lacks, particularly for:

• Complex multi-step task following
• Structured data extraction
• Maintaining persona consistency
• Following output format specifications exactly

For pure reasoning or coding tasks, the improvement over Llama is smaller — Qwen2.5-Coder and DeepSeek-Coder are better choices for those. See our best open-source coding models guide.