Build once.
Route anywhere.

Call the deterministic modelrelay/demo route without an account or key. Paid models are unavailable until project funding and usage accounting are enabled.

curl https://api.instamodel.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "modelrelay/demo",
    "messages": [{"role": "user", "content": "Confirm the gateway is alive"}],
    "stream": false
  }'

A gateway model ID currently maps to one configured upstream deployment. Paid publisher aliases must additionally snapshot the immutable revision, price, and terms before they are activated.

List models through GET /v1/models. Catalog metadata distinguishes platform-verified results from publisher-reported claims.

Instamodel-specific options live under the namespaced routing object, so standard OpenAI clients remain compatible.

{
  "model": "northstar-demo/clauseguard-32b",
  "routing": {
    "strategy": "balanced",
    "region": "us",
    "zero_retention": true,
    "allow_fallback": true
  }
}

Adapters translate protocols. The gateway owns routing, request IDs, usage records, canonical errors, and customer-facing SSE.

A provider is production-ready only after streaming, tool-call, Unicode, usage, cancellation, context-limit, and error-mapping conformance tests pass.

The browser flow submits to POST /api/publisher/models. Each submission creates a model, an immutable revision, a price version, and a commercial-terms version together. Its agreement digest commits the license identifier, artifact reference, runtime declaration, prices, royalty, and exact attestation version. Private-alpha organizations are quota-limited; validation, deployment, and alias activation remain operator approval gates.

1. 01DraftMetadata, license, base model, and task claims
2. 02QuarantinedArtifact present but not trusted or executable
3. 03ValidatingIntegrity, tokenizer, license, safety, smoke generation
4. 04DeployingApproved host receives digest-scoped artifact access
5. 05CanaryHealth, latency, output, and metering checks
6. 06LiveEligible for routing under the published alias

For large weights, create a multipart upload, send numbered parts of at most 50 MiB with an explicit Content-Length, then complete it with the returned ETags. Bytes land in a private quarantine bucket. Operators must explicitly enable direct uploads; defaults allow three active uploads, 50 GiB per artifact, and 100 GiB of declared artifacts per organization.

Clients declare size and SHA-256 before upload. Completion deletes and rejects an object whose actual size differs; an isolated validator must still independently hash and inspect the object before approval.

The paid-release ledger contract allows several provider attempts but exactly one final usage charge. Posted records are immutable and corrections use reversal entries. Automatic posting is still a release gate, so real-provider calls from console keys fail closed by default.

eligible_revenue = cash_backed_usage - refunds - fraud_reversals
owner_royalty  = eligible_revenue × royalty_bps / 10,000
host_payable   = measured_units × versioned_host_rate
platform_margin = eligible_revenue - owner_royalty - host_payable

Inference expense is kept separate from the model owner's percentage, so a routing change cannot silently alter creator economics.

Errors use the OpenAI object shape and every response carries x-request-id. Preserve that ID when opening a support issue or reconciling a usage line.

{
  "error": {
    "message": "No eligible deployment satisfies region=eu.",
    "type": "routing_error",
    "code": "no_eligible_deployment",
    "param": "routing.region"
  }
}