Model routing

A model id can express more than which weights you want. The grammar below is the one OpenRouter uses, so existing configuration works unchanged.

The grammar

author/slug                  cheapest healthy variant
author/slug:free             the rate-limited free variant, where published
author/slug:nitro            highest measured throughput
author/slug:floor            lowest price, ignoring speed
author/slug@provider         pinned to one provider
author/slug:free@infersia    both

With one provider in the catalogue these mostly resolve to the same endpoint. They exist now so that your code keeps working unchanged as the catalogue grows — a routing preference you set today will be honoured when there is something to choose between.

Aliases

Model ids are case-insensitive, and we accept the names you probably already have. All of these resolve to qwen/qwen3-8b:

qwen/qwen3-8b
Qwen/Qwen3-8B          ← the Hugging Face repo name
qwen3-8b
qwen3:8b               ← Ollama-style

If you mistype a model id, the error tells you the closest matches rather than just listing the catalogue.

Per-request routing controls

A provider object in the request body gives finer control. It is stripped before the request reaches the inference engine, so it never confuses an upstream.

request bodyjson

{
  "model": "qwen/qwen3-8b",
  "messages": [...],
  "provider": {
    "order": ["infersia"],
    "allow_fallbacks": true,
    "sort": "throughput",
    "quantizations": ["fp8", "fp16"],
    "data_collection": "deny"
  }
}

Parameter	Type	Description
`provider.order`	`string[]`	Provider slugs to try, in order.
`provider.ignore`	`string[]`	Providers to exclude entirely.
`provider.allow_fallbacks`	`boolean`	Whether to try other providers when the preferred one fails. Defaults to `true`.
`provider.sort`	`"price" \| "throughput" \| "latency"`	How to order candidates before applying the explicit order.
`provider.quantizations`	`string[]`	Only route to variants served at one of these precisions — for example `["fp8", "fp16"]` to refuse 4-bit entirely.
`provider.data_collection`	`"allow" \| "deny"`	Refuse providers that retain prompts.

The quantisation filter is the useful one. If you benchmarked a model at FP16 and your evaluation depends on that behaviour, pin it. A provider silently serving 4-bit is the single most common reason a model “got worse” without anyone changing anything.

How a variant is chosen

Only variants with at least one healthy endpoint are considered.
Free variants are never selected implicitly. A paying request cannot be silently downgraded onto a throttled endpoint because it sorted cheapest.
Within a variant, traffic is spread across endpoints weighted by their configured capacity, skipping any the gateway has recently seen fail.
If an endpoint fails before returning its first byte, we transparently try another. Once a stream has started we never retry, because splicing two generations together would produce output that is wrong in a way you can’t see.

Listing what’s available

curl https://api.infersia.com/v1/models

Unauthenticated, and it includes our extensions — quantization, context_length, supported_parameters — alongside the standard OpenAI fields. The official SDKs ignore the extra keys, so nothing breaks if you parse it with theirs.