API reference
Run our models — and models we build for you — behind one interface. Single calls return in-line for interactive use; large jobs go to a queue that streams results back as they finish. The same model, the same weights, the same answer either way, so what you validate in batch is what you serve in production.
REST server
https://api.belumind.comMCP server
https://mcp.belumind.com/inferenceEvery model below is callable through the same /v1/predict and /v1/batches endpoints — pass its id as model. Each one ships with a demo of the pipeline it runs.
noise-detect@2.1.0RealtimeBatchFinds interference buried in a signal and marks where it starts and stops, so downstream measurements can be trusted rather than silently corrupted. Tuned against the conditions it actually runs in, not a clean benchmark — the failure mode we care about is the noise nobody noticed.
ekyc-verify@1.4.2RealtimeReads an ID document, matches the face in front of the camera to the one on the card, and separates a live person from a photograph, a replay, or a generated face. Built for the phone the customer already owns, in the lighting they happen to be standing in.
csf-sign@1.0.0RealtimeBatchThe model from our CSF paper: a language-agnostic semantic layer that translates from any source language straight to sign language, without routing through a pivot language. Small enough to ship anywhere and fast enough to run on a CPU.
Every request carries a bearer token. Keys are scoped — an endpoint returns 403 if the calling key lacks the scope named in its reference entry, so an agent can be handed a key that reaches exactly one tool and nothing else.
MCP clients pass the same token as a bearer credential when connecting to the server URL above.
inference:runinference:batchAuthorization: Bearer blm_live_7f3c… Content-Type: application/json
/v1/predictinference:runSynchronous inference for interactive use. Returns in-line with a confidence score and the exact model version that produced the result, so the same call can be reproduced later.
predictmodelstringrequiredinputobjectrequiredoptionsobjectoptional200 OKSuccess. Body as shown in the example.400 Bad RequestThe request body failed validation. The response names the offending field.401 UnauthorizedMissing, malformed, or revoked API key.403 ForbiddenThe key is valid but lacks the scope this endpoint requires.429 Too Many RequestsRate limit exceeded. `Retry-After` carries the seconds to wait; usage is readable from /v1/usage.500 Internal Server ErrorSomething failed on our side. Requests are idempotent by `Idempotency-Key`, so a retry is safe.curl -X POST https://api.belumind.com/v1/predict \ -H "Authorization: Bearer $BELUMIND_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "noise-detect@2.1.0", "input": { "signal_url": "https://example.com/sample-recording.wav", "sample_rate": 48000 }, "options": { "threshold": 0.6 } }'
{
"prediction_id": "pred_5c7a13d9",
"model": "noise-detect@2.1.0",
"result": {
"detected": true,
"label": "broadband_interference",
"confidence": 0.93,
"segments": [
{
"start_ms": 1420,
"end_ms": 2180,
"confidence": 0.93
},
{
"start_ms": 7650,
"end_ms": 8010,
"confidence": 0.71
}
]
},
"latency_ms": 38
}POST https://api.belumind.com/v1/predictThe platform is in early access, so this console returns the documented example response rather than calling a live endpoint. Ask us for a key and the same console starts issuing real requests.
/v1/batchesinference:batchQueues inference over a large input set and returns immediately with a job id. Results stream back as each shard completes, so partial output is available before the whole job finishes.
create_batchmodelstringrequiredinput_urlstringrequiredcallback_urlstringoptionalmax_parallelismintegeroptional202 AcceptedJob accepted and queued.400 Bad RequestThe request body failed validation. The response names the offending field.401 UnauthorizedMissing, malformed, or revoked API key.403 ForbiddenThe key is valid but lacks the scope this endpoint requires.429 Too Many RequestsRate limit exceeded. `Retry-After` carries the seconds to wait; usage is readable from /v1/usage.500 Internal Server ErrorSomething failed on our side. Requests are idempotent by `Idempotency-Key`, so a retry is safe.curl -X POST https://api.belumind.com/v1/batches \ -H "Authorization: Bearer $BELUMIND_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "noise-detect@2.1.0", "input_url": "https://example.com/batch-inputs.jsonl", "callback_url": "https://example.com/hooks/batch-complete", "max_parallelism": 8 }'
{
"batch_id": "batch_71ad0e3f",
"status": "queued",
"total_items": 128450,
"created_at": "2026-08-31T09:14:22Z"
}POST https://api.belumind.com/v1/batchesThe platform is in early access, so this console returns the documented example response rather than calling a live endpoint. Ask us for a key and the same console starts issuing real requests.
/v1/batches/{batch_id}inference:batchReturns progress, per-shard state, and a results URL once output is available. Failed items are listed separately so they can be retried without re-running the whole job.
get_batch_statusbatch_idstringrequired200 OKSuccess. Body as shown in the example.400 Bad RequestThe request body failed validation. The response names the offending field.401 UnauthorizedMissing, malformed, or revoked API key.403 ForbiddenThe key is valid but lacks the scope this endpoint requires.429 Too Many RequestsRate limit exceeded. `Retry-After` carries the seconds to wait; usage is readable from /v1/usage.500 Internal Server ErrorSomething failed on our side. Requests are idempotent by `Idempotency-Key`, so a retry is safe.curl -X GET https://api.belumind.com/v1/batches/{batch_id} \ -H "Authorization: Bearer $BELUMIND_API_KEY"
{
"batch_id": "batch_71ad0e3f",
"status": "running",
"completed_items": 96200,
"failed_items": 12,
"total_items": 128450,
"results_url": "https://api.belumind.com/v1/batches/batch_71ad0e3f/results",
"estimated_completion": "2026-08-31T09:41:05Z"
}GET https://api.belumind.com/v1/batches/batch_71ad0e3fThe platform is in early access, so this console returns the documented example response rather than calling a live endpoint. Ask us for a key and the same console starts issuing real requests.
Errors use conventional HTTP status codes and always carry a machine-readable code alongside the human-readable message, so a client can branch on the former and log the latter. Every request is idempotent by Idempotency-Key, which makes retrying a 5xx safe.
{
"error": {
"code": "invalid_parameter",
"message": "`min_confidence` must be between 0 and 1.",
"param": "min_confidence",
"request_id": "req_4d91c0b7"
}
}The console returns documented examples while the platform is in early access. Tell us what you are building and we will get you a key.