#Hermes Agent Model Configuration

Hermes is model-agnostic: the same agent capabilities run on top of different LLMs. This chapter explains how to connect and switch models.

#Supported Providers

Also supported: NVIDIA NIM (Nemotron), Xiaomi MiMo, z.ai/GLM, Kimi/Moonshot, MiniMax, Hugging Face, and more.

#Choosing a Model Interactively

The simplest path:

hermes model

It lists available providers and models; follow the prompts to choose and enter credentials.

You can also switch mid-conversation:

/model openrouter:nous:hermes-3-405b

/model without arguments lists the options.

#Via the Config File

Model configuration ends up in ~/.hermes/config.yaml. A minimal example:

provider: openrouter
model: nous:hermes-3-405b

Or set it from the command line:

hermes config set provider openrouter
hermes config set model nous:hermes-3-405b

#Credentials & Environment Variables

Secrets like API keys can live in a .env file or your shell environment; Hermes reads them automatically:

# ~/.hermes/.env or your current shell
OPENAI_API_KEY=sk-...
ANTHROPIC_API_KEY=sk-ant-...
OPENROUTER_API_KEY=sk-or-...

Security tip: never commit keys to Git. Keep ~/.hermes/ private.

#Three Typical Setups

#1) Nous Portal (most convenient)

One-click OAuth login, plus an official tool gateway (web search, image generation, TTS, cloud browser) under one subscription:

hermes setup        # Choose Nous Portal and complete OAuth in the wizard
hermes portal info  # Inspect the current Portal config

#2) OpenRouter (most models)

A single API key unlocks 200+ models — great for comparison and cost optimization:

hermes config set provider openrouter
# Put OPENROUTER_API_KEY in .env
hermes model        # Pick a specific model

#3) Local vLLM (most private)

Deploy open models on your own GPUs so data and inference never leave your network:

provider: openai            # Use the OpenAI-compatible protocol
base_url: http://localhost:8000/v1
model: NousResearch/Hermes-3-Llama-3.1-405B

#Model Selection Guidance

Because Hermes ships 11 tool-call parsers, it tolerates different tool-call output formats well. Still, tool-heavy tasks benefit from models with strong tool-calling ability.

#Next Steps

• Tool System — let the model actually take action
• Memory System — let the model remember context
• Multi-Agent — different models for different tasks