1. The Short Answer: Model Choice Is Configuration
Hermes Agent doesn't ship its own mandatory model — it is model-agnostic. Which LLM you run is a setting you can change any time, not a decision you're locked into. You point the agent at a provider, it uses that model until you switch. That means the "right" model is whichever one clears two bars — a 64k-token minimum context window and dependable tool-calling — at a cost you're comfortable with. If you just want to get moving, run hermes setup --portal; if you already have an API key, run hermes model and paste it. For the bigger picture of what Hermes is, see our guide to Hermes Agent by Nous Research.
2. What Models Does Hermes Support?
Effectively any capable LLM, because Hermes talks to providers, not a single model. The supported endpoints are:
- • Nous Portal — Nous Research's hosted gateway to 300+ models. One subscription, many models.
- • OpenRouter — a single API key that fans out to many providers and models.
- • OpenAI — bring your own OpenAI API key.
- • Anthropic — bring your own Anthropic API key.
- • Any compatible endpoint — point Hermes at your own OpenAI-compatible URL (self-hosted or otherwise).
Because the model is decoupled from the agent, you can start on one provider and migrate later without rebuilding anything — the skills, memory, and config in ~/.hermes/ stay put.
3. How Do I Change the Model?
There are two commands you'll actually use. Run hermes model to choose (or switch) the provider and model and paste an API key. For beginners, hermes setup --portal is the smoothest on-ramp: it bundles a model together with four Tool Gateway tools — web search, image generation, text-to-speech, and a browser — so the agent is useful out of the box.
Your API keys are written to ~/.hermes/.env and the rest of your settings to ~/.hermes/config.yaml. Haven't installed Hermes yet? Start with our step-by-step Hermes Agent install guide and come back to pick a model.
4. What's the Minimum Context Window?
64,000 tokens. This is the one hard rule. Pick a model below that and your first messages will error out — Hermes needs the headroom for its system prompt, tool definitions, and working memory before you've even said anything substantial. So the very first filter on any candidate model is simple: does it advertise at least a 64k context window? If not, skip it. More context is generally better for long, tool-heavy sessions, but 64k is the floor you cannot go under.
5. How to Choose a Model: The Criteria Table
There's no single "best" model — there's the best fit for your task and budget. Work down these four criteria in order; the first one is a gate, the rest are trade-offs:
| Criterion | Why it matters | What to pick |
|---|---|---|
| Context window (≥64k) | Hard requirement — Hermes needs at least 64,000 tokens or your first messages will error out | Any model advertising 64k+ context; more headroom helps long sessions |
| Tool / function calling | Hermes is an agent — it lives or dies on reliable tool and function calls | A model with strong, well-tested tool-calling support |
| Cost per token | Cost is model-driven; compounding conversation history inflates spend over a session | A cheaper model for routine work; reserve premium models for hard tasks |
| Hosted API vs. your endpoint | Decides who runs inference — a managed gateway or infrastructure you control | Nous Portal for hosted convenience; your own key/endpoint for control |
6. Nous Portal vs. Bringing Your Own Key
Nous Portal is Nous Research's hosted gateway to 300+ models. It has subscription tiers — Free, Plus, Super, and Ultra — that bundle monthly credits, so you get access to a lot of models without wiring up individual provider accounts. It's the path of least resistance, and pairs naturally with hermes setup --portal.
If you'd rather bring your own API keys (OpenAI, Anthropic, OpenRouter, or a compatible endpoint), you self-host and pay only for your own inference — nothing on top. That's the better fit when you already have provider accounts, want direct control over which exact model runs, or need to keep inference on infrastructure you manage.
7. Keeping Cost Under Control
A useful mental model: cost is model-driven, not agent-driven. Hermes itself is free and open source — what you pay for is inference, and that's set by the model you chose. The subtle part is that conversation history compounds: every turn, the agent re-sends prior context, so a long session quietly inflates token spend even if each message looks small.
Two general levers help. First, run a cheaper model for routine work and reserve pricier ones for genuinely hard tasks. Second, use history compaction to trim the context the agent carries forward. Both attack the same thing — the number of tokens flowing through the model — which is where the money goes.
8. A Simple Way to Decide
- • Just starting out? Run
hermes setup --portal. You get a working model plus web search, image generation, TTS, and a browser without touching any provider dashboards. - • Already have an API key? Run
hermes model, pick your provider, paste the key, and confirm the model clears 64k context. - • Cost-sensitive or high-volume? Default to a cheaper model, turn on history compaction, and only reach for a premium model when a task actually needs it.
- • Need control over inference? Bring your own endpoint and self-host — you pay only for your own inference.
9. Frequently Asked Questions
What's the best model for Hermes Agent?
There isn't one universal best — Hermes is model-agnostic, so the best model is the one that fits your task and budget. Filter for a 64k+ context window and strong tool-calling first, then balance cost. Beginners can let hermes setup --portal pick a sensible default.
Can I switch models later?
Yes. Because model choice is configuration, run hermes model any time to switch providers or models. Your skills, memory, and settings in ~/.hermes/ are unaffected.
Can I use my own local or self-hosted model?
Yes — point Hermes at any compatible endpoint with hermes model. As long as it exposes an OpenAI-compatible API and meets the 64k-token context minimum, Hermes can drive it, and you pay only for your own inference.
Do I have to pay for Nous Portal?
Nous Portal has a Free tier alongside paid Plus, Super, and Ultra tiers that bundle monthly credits. If you'd rather not use Portal at all, bring your own API keys and pay only your own provider for inference.