Hermes Agent & Nous Research

Who Is Nous Research?

Nous Research is an AI research collective focused on open, decentralized AI. Their public output spans three categories:

• Open-weight models — the Hermes series (Hermes 3, Hermes 4), built on top of base models from Meta, Mistral, and others.
• Research papers — on distributed training, model alignment, and inference efficiency.
• Open-source tools — including Hermes Agent, the autonomous-agent framework this guide is about.

Their stated mission is to make capable AI accessible outside the centralized lab ecosystem — everything they ship is open-source or open-weight.

The Hermes Model Series

The Hermes models are Nous's flagship line. They're fine-tunes of strong open base models, optimized for instruction-following, tool use, and long-context reasoning.

• Hermes 3 — 8B, 70B, and 405B variants based on Llama 3.1. Strong tool use and structured output.
• Hermes 4 — the next-gen series with improved reasoning and broader tool catalogs.

You can pull the weights from Hugging Face and run them locally via Ollama, llama.cpp, vLLM, or LM Studio. Hermes Agent ships with first-class adapters for all four runtimes.

How Hermes Agent Fits

Hermes Agent is the “application layer” that turns a Hermes model into a usable autonomous agent. It handles:

• Multi-channel deployment (Telegram, Discord, WhatsApp, web)
• Multi-level persistent memory across sessions
• Tool calling and MCP server integration
• Authentication, rate limiting, and session management

You don't have to use a Hermes model with Hermes Agent — the framework is model-agnostic and works with Anthropic, OpenAI, Google, OpenRouter, and any OpenAI-compatible endpoint. But the framework is designed to make Hermes models first-class, with tuning for their tool-use format and prompt conventions.

Why “Hermes”?

Hermes is the Greek god of messengers, boundaries, and translation between worlds — a fitting name for an agent that lives at the boundary between humans and models, and translates instructions across chat platforms, tool calls, and memory layers.

Open-Source License

Hermes Agent is MIT-licensed. The Hermes model weights are released under various open licenses depending on the base model (Llama community license for the Llama variants, Apache 2.0 for the Mistral variants). Check the model card on Hugging Face for the exact terms before commercial use.

The Bigger Picture: Distributed AI

Nous's longer-term bet is that AI infrastructure should be decentralized: open weights, local inference, user-owned memory. Hermes Agent is part of that bet — it's the runtime that lets you keep your AI assistant on your own hardware, with your own memory store, talking to your own platforms.

How to Run Hermes Models with Hermes Agent

The cleanest local setup uses Ollama for serving and Hermes Agent for the runtime:

# Pull a Hermes model via Ollama
ollama pull hermes3:8b

# Point Hermes Agent at the local Ollama endpoint
{
  "models": {
    "providers": {
      "ollama": {
        "baseUrl": "http://localhost:11434/v1",
        "apiKey": "ollama"
      }
    }
  },
  "agents": {
    "defaults": {
      "model": { "primary": "ollama/hermes3:8b" }
    }
  }
}

See the Hermes + Ollama guide for the full local-inference walkthrough, and the Hermes + Llama guide for the broader llama.cpp / vLLM paths.

Where to Follow Nous

• nousresearch.com — blog and announcements
• github.com/NousResearch — open-source repos including Hermes Agent
• huggingface.co/NousResearch — model releases
• @NousResearch on X — release notes and research threads

Try Hermes Without Setup