Hermes Agent Guide
Best Models for Hermes Agent: A Method, Not a Leaderboard
Hermes talks to more than forty model providers and switching between them is one command with no code changes. That changes the question. "Which model is best?" matters far less when a wrong answer costs you thirty seconds — so this guide is about how to choose for a given job, and what Hermes gives you to choose with.
Start with the picker
Everything below assumes one command:
hermes modelIt lists what your configured credentials can actually reach and saves your choice as the default. This is worth internalising before reading any model recommendation, including the ones here: on Hermes, model selection is a reversible, thirty-second decision. Treat recommendations as starting points to test, not commitments.
What Hermes can reach
The provider list is unusually broad, and it spans several different kinds of access, which matters more than the raw count:
| Access type | Examples | Set up with |
|---|---|---|
| Free tier (key, no payment) | OpenRouter free models | An OpenRouter key in ~/.hermes/.env |
| API key | OpenRouter, DeepSeek, z.ai (GLM), MiniMax, Groq, Fireworks, NVIDIA, Hugging Face | A key in ~/.hermes/.env |
| OAuth / subscription | Nous Portal, Anthropic, GitHub Copilot, Qwen, MiniMax, xAI Grok | hermes model browser login |
| Cloud platform | AWS Bedrock, Azure AI Foundry, Google Vertex AI | hermes model with platform credentials |
| Local / self-hosted | LM Studio, Ollama, vLLM | Picker or a base URL |
| Anything OpenAI-compatible | Together AI, your own gateway | Custom endpoint |
Choosing by job, not by benchmark
An agent is not a chatbot, and the qualities that matter are not the ones leaderboards rank. Three properties dominate in practice.
Tool-calling reliability
Hermes works by calling tools in a loop. A model that writes beautiful prose but emits malformed tool calls will fail at the only thing that matters here, and it will fail intermittently, which is worse than failing outright. If you test one thing before committing, test whether the model calls tools correctly under your actual prompts.
Cost per loop, not per token
Agent turns are multi-step: a single request can involve several model calls as tools run and results come back. A model that is marginally cheaper per token but needs twice the steps is not cheaper. Judge by what a representative task costs end to end.
Latency you will actually feel
For an agent you message and walk away from, latency barely matters. For one you converse with, it is most of the experience. Match the model to how you interact — and note you can change it per situation rather than picking one forever.
A practical way to start
- Start with a free-tier model on OpenRouter to confirm your Hermes install works end to end, before any billing relationship exists. Skip
opencode-free: it only worked inside OpenCode's own client when we tested it in September 2026 and refused Hermes. - Use that same aggregator key to compare several paid models too, without opening accounts with each vendor.
- Run your real workload, not a benchmark prompt. Watch for malformed tool calls and step counts.
- Once a model earns its place, go direct to that vendor if the economics justify it. See the BYOK guide for using your own keys.
- Revisit occasionally. The catalogue moves quickly and switching is still one command.
Why this beats a recommendation list: any specific model named here will be stale within months, while the method survives. Hermes is built so that being wrong about a model is cheap — use that rather than trying to be right in advance.
Keeping the choice honest
- Keep credentials in
~/.hermes/.env, never in a Git-tracked file. - Use a separate key per environment so you can revoke one without disrupting the other.
- Watch spend at the provider for the first few days after switching; agent loops multiply usage.
- Re-test tool calling after any model change, not just after a provider change.
The same question for the other framework is covered in Best models for OpenClaw.
Hermes Agent model selection FAQ
How do I change the model Hermes uses?
Run hermes model. It is an interactive picker covering every configured provider, and your selection is saved as the default. There are no code changes involved, which is the single most useful fact about model choice on Hermes — a wrong pick costs you one command, not a migration.
Can I try Hermes without paying for an API key?
Mostly, but not through the opencode-free provider. Hermes lists it as keyless, yet OpenCode's free tier only answers requests that come from OpenCode's own client. From anywhere else, including Hermes, it has been refusing requests (HTTP 403, "free tier can only be used from within OpenCode", as tested in September 2026), so you can end up with no replies at all. A working zero-cost start is an OpenRouter account with a free-tier model, which needs a key but no payment. Free-tier models are rate limited and change often.
Is there one best model for Hermes?
No, and any page claiming otherwise is selling something. The useful question is which model for which job: an agent that runs long tool-calling loops has different needs from one summarising a document. Because switching is one command, the practical method is to test on your own workload rather than trusting a leaderboard.
Does Hermes lock me into one provider?
No. It supports 40-plus providers — OpenRouter, Anthropic, Google, DeepSeek, z.ai, MiniMax, Groq, Fireworks, NVIDIA, AWS Bedrock, Azure, Vertex, Hugging Face, LM Studio, Nous Portal and more, plus a custom endpoint for anything OpenAI-compatible. Credentials live in ~/.hermes/.env and are yours.