Limitations | Sarvam API Docs

Know these before you build — most are inherent to how SageMaker or the model container works, and each has a documented path around it.

Network & connectivity

No outbound internet. Endpoints run with network isolation on, so the container cannot make outbound calls — no webhooks or callbacks originate from the model. Your application polls or reads results itself (from the invoke response, or from S3 for async/batch).

Duration & size

Limit	Applies to	Work around it with
30 s max audio	Saaras v3 real-time REST	Streaming (bidirectional, for longer audio)
8 or 16 kHz, mono 16-bit PCM	Saaras v3 streaming	Resample before streaming
2500 chars real-time / 3500 chars streaming	Bulbul v3 text input	Split long text, or use SSE / bidirectional
≤ 24000 Hz output	Bulbul v3 `speech_sample_rate`	Request `24000`; higher is rejected (`400`), not upsampled
5 pages per request (recommended)	Sarvam Vision sync	Async or batch for larger documents
500 pages per document (hard cap → `413`)	Sarvam Vision	Split the document; route bulk work to async or batch
50 MB per file	Sarvam Vision async	Split very large documents
60 s max	SageMaker real-time `InvokeEndpoint` — all sync endpoints	Streaming (STT/TTS) or async / batch (Vision)
~6 MB request/response	All real-time endpoints	Async (large payloads) or Batch

The 60-second and ~6 MB ceilings are SageMaker real-time InvokeEndpoint limits, not Sarvam limits — async and batch handle larger inputs and longer jobs. A Sarvam Vision sync request should stay to 5 pages or fewer so it finishes inside 60 s; send multi-page or bursty document work to async (up to 50 MB per file) or batch.

Text-to-Speech sustained-overload latency. The current Bulbul v3 shipping config runs a single web worker (WEB_CONCURRENCY=1), so under sustained overload served p95 drifts upward and throughput drops rather than staying flat. Size for headroom and back off on 503 service_overloaded; do not rely on flat latency once the endpoint is saturated. See Configure & tune.

Packaging & regions

Region-specific ARNs. A model package ARN only works in the region it was issued for. Subscribe and copy the ARN for each region you deploy into.
Version split. The package version (e.g. saaras:v3.1) is not the request-time model id (e.g. saaras:v3). See the API reference.

Observability

Standard SageMaker CloudWatch metrics and container logs are available (see Operations). Custom in-container business metrics are not emitted.

Hitting a limit that blocks your use case? Tell us at developer@sarvam.ai — several of these have async/batch paths, and we can advise on sizing.