Hermes by Nous Research

Hermes is a family of open models from Nous Research, fine-tuned specifically for function calling, tool use, and agentic behavior. Versions include Hermes 2, Hermes 3, and Hermes 4.

Why it matters

Most base models are tuned for chat. Hermes is tuned to call tools correctly — to output structured JSON when asked, and to follow a rigid function-calling format. That makes it a strong open-source choice for local agents.

Run it locally

ollama run hermes3

Or pull the exact tag you want:

ollama pull nousresearch/hermes-3-llama-3.1-70b

Use Hermes as an agent backend

Hermes supports the standard chat-completions tools schema, so any client that speaks OpenAI-compatible APIs can drive it:

{
  "model": "hermes3",
  "messages": [{"role": "user", "content": "What is 24 * 7?"}],
  "tools": [{
    "type": "function",
    "function": {
      "name": "multiply",
      "description": "Multiply two numbers",
      "parameters": {
        "type": "object",
        "properties": {
          "a": {"type": "number"},
          "b": {"type": "number"}
        },
        "required": ["a", "b"]
      }
    }
  }]
}

A tool-tuned model like Hermes will reply with a structured tool_calls object instead of prose — exactly what an agent harness needs.

The bigger lesson

"Open models" aren't all the same. Agentic ability is a fine-tuning target, and Hermes is the clearest example: same base model, dramatically better tool use.

When picking a local model for an agent, look for "tool calling" or "function calling" in the model card — not just benchmark scores.