Sparround

Model providers and Nous Portal

Hermes is model-agnostic: the agent is one thing, the provider that serves the model is another. Your first decision after installing is which provider to connect — Nous Portal, OpenRouter, Anthropic, OpenAI, Google, DeepSeek, xAI, a local Ollama and others are supported.

You can configure several providers at once and switch between them inside a session with /model.

RouteWhen to pick itWhat to watch
Nous PortalWhen you want many models behind one subscription; it is the route the docs recommendA subscription model — check the price on the portal site
Direct provider (Anthropic, OpenAI, ...)When you already have a corporate account and keyA separate key and quota per provider
OpenRouterTo try many models behind one keyIt is an intermediary — requests pass through it
Local model (Ollama and similar)When data must not leave the machineContext and speed limits; needs capable hardware

One hard requirement: the docs call for a model with at least 64K tokens of context. The reason follows from the architecture — the system prompt, tool schemas, skill list and conversation history already fill a large part of the window. A small-context model may technically work, but it drops straight into compression and quality suffers.

Two extra mechanisms exist:

  • Provider routing — decide exactly which provider handles which work (an expensive model for the main task, a cheap one for auxiliary jobs)
  • Fallback providers — automatically move to the next provider when one hits a rate limit

Auxiliary work is configured separately: compression, image analysis, title generation. Routing those to a cheap, fast model is the typical first optimisation — it barely affects output quality and noticeably affects the bill.

Practice. Configure two providers (one main, one cheap auxiliary) and point auxiliary.compression at the cheap one. Done means: hermes config get auxiliary.compression.model returns what you expect, and after one long session hermes insights shows the two models' costs separately.

📚 Sources and documentation