> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hq.zone/llms.txt
> Use this file to discover all available pages before exploring further.

# Models & inference

> How HQ routes every agent call across the frontier and serves the sensitive, high-volume work on its own GPUs — one model layer beneath qOS.

Decision intelligence needs more than one model. **Agent Inference** is the model layer beneath HQ: every agent reaches frontier models through one interface, routed to whichever fits the task, alongside open models HQ serves on its own GPUs.

This page is part of [qOS](/concepts/qos) and complements [How HQ works](/concepts/how-hq-works).

## Multi-model routing

Every agent reaches frontier models from **Anthropic, OpenAI, xAI, and Cerebras** behind one interface. Each call goes to the model that fits the task — quality, latency, or cost — so the platform follows the frontier as it moves rather than being wired to whichever vendor was integrated first.

Picking a model is part of defining an agent, not a buried platform setting. Give a research agent a frontier reasoning model and a triage agent a fast, small one, and run them side by side in the same workspace.

<CardGroup cols={2}>
  <Card title="A model per agent">
    Give each agent the model that fits its job, and run a fleet where every member is on a different model at once.
  </Card>

  <Card title="Automatic failover">
    If a provider errors or goes down, the call retries against a healthy one inside the same request.
  </Card>
</CardGroup>

→ See what's available via [List models](/api-reference/agents/list-models), and select one per agent.

## Own-GPU inference for sensitive work

There is a quieter problem underneath model choice: most agent platforms ship your data to a third-party API just to embed or rerank it — the highest-volume calls they make. HQ answers this by serving the cheap, constant, and data-sensitive paths on hardware it runs itself.

<CardGroup cols={2}>
  <Card title="Open models we serve ourselves">
    Open-weight models run on our own GPUs, so the high-volume and data-sensitive paths never have to leave our infrastructure.
  </Card>

  <Card title="Embeddings, in-house">
    The vectors behind memory and search come from an embedding model we host, so indexed text is never sent to an outside API to be turned into numbers.
  </Card>

  <Card title="Reranking, in-house">
    A self-hosted reranking model does the final relevance pass on memory and corpus lookups, with no third-party round-trip.
  </Card>

  <Card title="A batch path at fleet scale">
    Large non-interactive jobs run through a batch lane for work that can wait rather than answer in the moment.
  </Card>
</CardGroup>

The split is deliberate: frontier models for the hard, rare reasoning; owned open models for the cheap, constant, and sensitive work — not a degraded fallback.

## Independent of the runtime

The model and the runtime are two separate dials. Each agent runs on a vendor's official agent loop — the Claude Agent SDK, the OpenAI Agents SDK, or the open OpenCode runtime — and any model can run under any of those runtimes, so each choice is made on its own merits.

<Note>
  The model layer sits inside qOS next to search, cache, and storage, sharing the same hardware, isolation, and audit trail as everything else an agent touches. Adding or swapping a model is a config change, with nothing to alter in the agents on top.
</Note>