Quickstart

1. Choose a project and obtain a key

Log in to Console, choose an authorized workspace and project, and open API keys. Use an inference key with inference:use. Your role must permit key creation; ask your workspace administrator if it does not.

Keep the key in your client environment. Never paste it into documentation URLs, source control, or a public browser application.

export INFERMAX_API_KEY="<your-project-inference-key>"
export INFERMAX_API_BASE_URL="https://api.infermax.ai"

2. Discover your entitled models

curl "$INFERMAX_API_BASE_URL/v1/models" \
  -H "Authorization: Bearer $INFERMAX_API_KEY"

Choose a model ID from the returned data array. Listing a model does not guarantee it is ready. Check GET /v1/models/{id}/health before sending work.

3. Send a Chat Completion

Replace <entitled-model-id> with an ID from your project’s catalog. The message below is a synthetic example.

curl "$INFERMAX_API_BASE_URL/v1/chat/completions" \
  -H "Authorization: Bearer $INFERMAX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"<entitled-model-id>","messages":[{"role":"user","content":"Say hello in one sentence."}],"max_completion_tokens":64,"stream":false}'

Use stream: true and curl -N for Server-Sent Events. Tool execution belongs to your client. Model capabilities and request limits depend on the entitled, qualified deployment.

See the request schema, response formats, and cancellation contract →

4. Handle uncertainty explicitly

An optional Idempotency-Key identifies a Chat Completion attempt. Preserve the same key and body when checking a retry. Without a key, a new request may create another attempt after a timeout. Do not automatically resend work when dispatch is uncertain.

For Compute Access, begin with eligible offers, then reserve capacity and create an instance using the required idempotency headers. Inference-only keys cannot manage compute. Read the Compute Access reference before provisioning.