> For clean Markdown of any page, append `.md` to the page URL.
> For a complete documentation index, see https://docs.sarvam.ai/llms.txt.
> For full documentation content in one file, see https://docs.sarvam.ai/llms-full.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.sarvam.ai/_mcp/server.

# How it works

> The architecture of a Sarvam self-hosted deployment: AWS Marketplace entitlement, a SageMaker model package, and an inference endpoint running in your own VPC. Plus how to choose real-time, async, or batch.

A self-hosted deployment has three pieces: an **AWS Marketplace subscription** (entitlement), the **Sarvam model package** (the pre-built container Sarvam publishes), and a **SageMaker endpoint** you create from it inside your own account.

```mermaid
flowchart LR
    subgraph MP["AWS Marketplace"]
        L["Sarvam listing — model package<br />Saaras v3 STT · Bulbul v3 TTS · Sarvam Vision"]
    end
    subgraph ACC["Your AWS account and VPC — network-isolated"]
        subgraph SM["Amazon SageMaker"]
            C["Sarvam model container<br />GPU instance"]
        end
        APP["Your application"]
    end
    L -->|"subscribe (entitlement)"| C
    APP <-->|"InvokeEndpoint · your data stays in-account"| C
```

* **The model package** is a versioned container Sarvam maintains. You never build or patch it — you subscribe to a version and deploy it.
* **The endpoint** runs on GPU instances you choose, in your region and VPC. It is network-isolated: no outbound internet from the container.
* **Your data** (audio, documents) is sent only to your own endpoint and stays in your cloud.

## Deployment modes

#### Real-time

An always-on endpoint that returns results in the API response. For interactive and synchronous workloads.

#### Async

Submit a job, get results from S3 when ready. Supports large payloads and **scale-to-zero** when idle.

#### Batch transform

Process a whole S3 prefix of files in one job, then shut the instances down. For bulk/offline processing.

Not every model supports every mode:

* **Saaras v3 (STT):** **real-time** (audio ≤ 30 s) and **streaming** (bidirectional SigV4 HTTP/2, for longer or continuous audio). Batch is not supported.
* **Bulbul v3 (TTS):** **real-time**, **server-side streaming (SSE)**, **bidirectional** (SigV4 HTTP/2), and **batch transform**.
* **Sarvam Vision:** **real-time** (sync), **async**, and **batch transform**.

### Which mode should I use?

| Your workload                                                 | Use                                              |
| ------------------------------------------------------------- | ------------------------------------------------ |
| Live captions, interactive transcription, short audio clips   | **Real-time** (Saaras v3)                        |
| Long or continuous audio                                      | **Streaming** (Saaras v3)                        |
| Short text prompts, interactive speech                        | **Real-time** (Bulbul v3)                        |
| Low-latency playback or long text (≤ 3500 chars)              | **Streaming — SSE or bidirectional** (Bulbul v3) |
| A small document that finishes within 60 s, interactively     | **Real-time** (Vision)                           |
| Large documents, occasional or bursty traffic, cost-sensitive | **Async** (Vision — scale-to-zero)               |
| Bulk/offline processing of many documents in S3               | **Batch transform** (Vision)                     |

> **Note**
>
> For Saaras v3, the real-time path accepts audio up to **30 seconds**; for longer or continuous audio use the **streaming** endpoint. See [Deploy Speech-to-Text](/api/self-hosted/sagemaker/deploy-saaras).

> The architecture of a Sarvam self-hosted deployment: AWS Marketplace entitlement, a SageMaker model package, and an inference endpoint running in your own VPC. Plus how to choose real-time, async, or batch.