Skip to main content

​
Routers

Routing is configured on the Routers page. A router holds three size tiers plus per-task-group assignments:
  • Big: serves clearly hard requests.
  • Medium: the default; serves the bulk.
  • Small (optional): serves quick machine tasks.
The cheaper tiers only ever serve machine-shaped work: classification, extraction, summarization, strict output formats, template-style prompts, and tool loops. When a person is talking to the model, the response comes from the top of the router. To make the decision in your own code without executing the call, use POST /api/v1/route. A project can hold several routers:
  • Exactly one is the auto router. It serves calls that say model: "auto" or name no model. The others are drafts: stage a new setup there and click Make auto to switch, instantly and reversibly. The auto router cannot be deleted, only replaced.
  • The first inference call that names a real model creates a router anchored on that model, listed as “From your code” and editable like any other. The anchored model is the router’s ceiling and answers every conversational call; its Medium and Small tiers take only the machine-shaped tasks underneath. Deleting one means the next call naming that model recreates it.
Every new project starts with a Default router: GLM 5.2 serves the bulk (Medium), clearly hard requests escalate to Claude Fable 5 (Big), and quick machine tasks take Qwen 3.5 9B (Small). A project created with Chinese models off gets Grok 4.6 as Medium and Llama 3.2 3B as Small instead.

​
Task groups

Each request is matched to your project’s task groups. A task group you assigned a model to serves that model (tier group); task groups with no assignment fall to the tier defaults above. Pass your session id with each call and a conversation stays on the model of its first decision for up to 24 hours (tier pinned); see session pinning.

​
Smart routing

Smart routing (under Project Permissions, on by default) is whether the router may pick the serving model at all. On is everything this page describes: tier escalation, the Small tier, task-group routes, record picks, session pins, and routing through a named model’s anchored router. Turn it off and you only ever get responses from the model you select: a request that names a model is served exactly that model, and auto is served by the auto router’s Medium model, deterministically (tier fixed). Your routers, records and recommendations stay intact for when you turn it back on.

​
Automatic updates

The Routers page shows each task group’s observed success rate per model, that model’s newest score in the task group’s generated environment when one exists, and a recommendation: the cheapest model whose recorded results match the best one. Automatic updates (under Project Permissions, on by default) decides what happens to that recommendation:
  • On: for task groups you have not assigned a model to, Belvedir promotes the recommended model, and machine-shaped requests in that task group serve it live (tier group-record). Conversation never routes this way, task groups with too few results keep their tier default, and a model you assigned yourself always wins over the record.
  • Off: nothing switches by itself. The router editor shows a Recommended band for each task group with its observed record, and an Approve button puts that one assignment into effect.
Your fine-tuned models earn traffic the same way, by winning on the record, once registered under Cloud Inference. The request and response contract, headers included, is on chat completions.