Welcome to Model APIs Documentation

Model APIs bring two model families together. Both use the same API key, the same sarvamai SDK, and the same credit balance, so you can mix them in one app.

	Token Factory (Open-Weight)	Sarvam Models
What it is	Open-weight LLMs served on Sarvam infrastructure	Sarvam’s own models for Indian-language speech, translation, and documents
Models	Sarvam 105B, GLM-5.3, Gemma 4, DeepSeek, Kimi	Saaras, Bulbul, Mayura, Sarvam-Translate, Sarvam Vision
APIs	Chat Completions and Responses on `/v2`, compatible with the OpenAI SDK	Speech to Text, Text to Speech, Text Processing, Document Intelligence, Document Translation, Dubbing, Voice Cloning
Use it to	Reason, write code, call tools, and run agents	Hear, speak, translate, and read Hindi, Tamil, Telugu, and every major Indian language, including code-mixed input
Start here	Token Factory overview	Speech to Text or Text to Speech

Get started

Quickstart (< 5 minutes)

Get your API key and make your first request, with code for cURL, Python, and JavaScript.

SDKs & Libraries

Official Python and JavaScript clients, with async, retries, and streaming built in.

Pricing, Credits & Rate Limits

Rates for every model in Indian Rupees, how credits work, and the limits on each plan.

API Reference

REST and WebSocket specs for every endpoint, with request and response schemas.

Token Factory (Open-Weight)

Open-weight LLMs on https://api.sarvam.ai/v2, compatible with the OpenAI SDK. Switch models by changing the model field.

Model	Model ID	Best for
Sarvam 105B	`sarvam-105b`	Indian-language reasoning, coding, and agents
Sarvam 105B Conversations	`sarvam-105b-conversations`	Real-time dialogue and voice agents
GLM-5.3	`glm5.3`	Coding and long-horizon agents
GLM-5.3 Flash	`glm5.3-flash`	Fast reasoning with image input
Gemma 4 31B	`gemma4`	Images and low-latency answers
DeepSeek V4.1 Flash 1017	`deepseekv4.1-flash-1017`	Long documents and cost-sensitive reasoning

DeepSeek V4 Pro, Kimi K3.1 Pro, Qwen 3.8 27B, the Clef decision model, and EmbeddingGemma 2 text embeddings are on Provisioned Throughput and are enabled per organisation.

Token Factory overview

What Token Factory is, and the two-line switch from the OpenAI SDK.

Compare models

Context windows, modalities, reasoning, and feature support side by side.

Chat Completions

Streaming, tool calling, reasoning, and structured output.

Responses API

Stateless responses with reasoning, structured output, and tool calling.

Moving from another provider? Follow the migration guide for OpenRouter, Fireworks AI, or Together AI.

Sarvam Models

Speech to Text

Transcribe and translate audio in real time or in batch with Saaras, including code-mixed speech.

Text to Speech

Generate natural, expressive speech in Indian languages with Bulbul, streamed or in one request.

Text Processing

Translate, transliterate, and detect languages with Mayura and Sarvam-Translate.

Document Intelligence

Turn PDFs, scans, and handwritten documents into structured data with Sarvam Vision.

Document Translation

Translate whole PDF, Word, spreadsheet, presentation, and web page files into 23 languages, layout preserved.

Dubbing

Localize video and audio into Indian languages, keeping each speaker’s voice.

Instant Voice Cloning

Generate speech in a cloned voice from a short reference clip, in 12 Indian languages.

Self-Hosted Deployments

Run Saaras, Bulbul, and Sarvam Vision in your own AWS account on Amazon SageMaker.

Combine them

Both families use the same key, so a voice agent is three calls: Saaras hears the caller, an LLM decides what to say, and Bulbul says it.

Build a voice agent with LiveKit

Wire Saaras, Sarvam 105B, and Bulbul into a real-time voice pipeline.

Build a voice agent with Pipecat

The same pipeline on Pipecat.

Build with your tools

Integrations: LiveKit, Pipecat, Twilio, Exotel, Vobiz, Vapi, n8n, the Vercel AI SDK, and LangChain.
Developer Tools: the MCP server, Markdown and llms.txt, Context7, and Agent Skills for AI coding assistants.
Cookbook: step-by-step notebooks, from first calls to full pipelines.
Migrations: switch speech and Gemini workloads from ElevenLabs, Cartesia, Deepgram, or Gemini.

Resources

All Models: every model across both families, with language support and how to choose.
Building for Indian Languages: language coverage, code-mixing, scripts, numerals, and telephony audio.
Authentication: send your API subscription key, or a Bearer token, with every request.
Errors & Troubleshooting: error codes, retry strategies, and fixes for common issues.
Account & Billing: organisations, workspaces, team invites, and payments.
Changelog: new features, model updates, and breaking changes.
Talk to us: reach the team for bugs, feature requests, and enterprise questions.
Discord: early access to releases, office hours with the team, and India’s AI builder community.