> 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.

# Self-Hosted Deployments

> Deploy Sarvam's speech and document-intelligence models inside your own AWS account with Amazon SageMaker. Your audio and documents never leave your VPC.

Run Sarvam's models **inside your own cloud**. Instead of calling the Managed API at `api.sarvam.ai`, you subscribe to a Sarvam model package on the **AWS Marketplace** and deploy it as an Amazon SageMaker endpoint in your own account and VPC. Your audio, documents, and transcripts never leave your infrastructure.

Self-hosted deployments are a different delivery model from the [Managed Sarvam API](/api/getting-started/welcome). Same models, two ways to run them — see [Managed vs Self-Hosted](/api/self-hosted/hosted-vs-self-hosted) to choose.

## Models available for self-hosting

#### [Saaras v3 — Speech-to-Text](/api/self-hosted/sagemaker/deploy-saaras)

Indic-first ASR across 23 languages with five output modes (transcribe, translate, verbatim, translit, codemix). Real-time and streaming.

#### [Bulbul v3 — Text-to-Speech](/api/self-hosted/sagemaker/deploy-bulbul)

Indic-first TTS with 38 voices. Text in, base64 or streamed audio out — real-time, server-side streaming (SSE), and bidirectional.

#### [Sarvam Vision — Document Intelligence](/api/self-hosted/sagemaker/deploy-vision)

3B-parameter vision-language model for OCR and document parsing across 23 languages. PDFs and images in, clean HTML or Markdown out.

## Why self-host

#### Data stays in your cloud

Audio and documents are processed inside your own VPC. Nothing is sent to Sarvam — ideal for BFSI, healthcare, and government workloads with data-residency requirements.

#### Predictable, usage-independent cost

You pay AWS for the instance hours you run — a flat software fee plus infrastructure — instead of a per-call price.

#### Low, in-network latency

The endpoint lives next to your application in the same region and VPC, with no public-internet round trip.

#### Runs in air-gapped setups

Endpoints deploy with network isolation enabled — the model container has no outbound internet access.

## How it works, end to end

#### Subscribe on AWS Marketplace

Accept the EULA on the Sarvam listing and copy the model package ARN for your region.

#### Deploy a SageMaker endpoint

Create a model from the package, then a real-time, async, or batch endpoint — via the console, boto3, or Terraform.

#### Invoke it from your app

Call `InvokeEndpoint` with your audio, text, or document. Results come back as JSON (STT), audio (TTS), or HTML/Markdown (Vision).

There is **no license key or license proxy** to run. AWS Marketplace handles entitlement — once you're subscribed, the endpoint is fully licensed for the instances you run it on.

## Next steps

#### [Get started on SageMaker](/api/self-hosted/sagemaker/get-started)

Prerequisites, IAM setup, subscribing, and quotas — everything before your first deploy.

#### [API reference](/api/self-hosted/sagemaker/api-saaras)

The exact `InvokeEndpoint` request and response contract for each model.