Serverless Endpoints
Call models over HTTPS with pooled capacity or a dedicated provider endpoint.
Pooled endpoint
One model, routed to whichever qualified provider node is free. No capacity or latency guarantee during the pilot.
Supported API and billing parameters
Use POST /api/v1/chat/completions with your NexusCompute bearer API key and a unique Idempotency-Key header. Send model, messages (system/user/assistant roles with text content), and max_tokens (1–1024; default 256). Sampling is fixed at temperature 0.
Streaming, tools, images, stop sequences and other request parameters are not supported and return a validation error. Capacity failures return 503. After a response timeout, reuse the same idempotency key; do not submit a new paid request. A repeated key returns the request metadata and status, never a regenerated output.
Content handling: there is no persistent content history. Inputs are kept in temporary memory for at most 120s; completed outputs for at most 60s or until the client acknowledges receipt (POST /api/pool/requests/:id/ack), whichever is first. There is no durable result recovery, so read the output when it is delivered. A shared Redis transport for multi-instance operation exists but is disabled by default until verified and is not part of current live capacity.
Input is metered over the submitted JSON messages and output with the model’s registered server tokenizer, not a node-reported token count. Charges include input and output, rounded up to whole cents. The maximum charge is reserved from your prepaid balance before dispatch; failed or expired work releases the hold.
Private request transport is not ready, so new prompts are not accepted. Existing request IDs can still be checked.