AD
Sarvam AI logo

Sarvam AI

Freemium🇮🇳High Burn

India's sovereign full-stack AI platform for all Indian languages

83

Overall score

36

Heat score

Pricing

Free TrialFree credits on signup
Pay-As-You-GoUsage-based (no monthly fee)
BusinessCustom plan with bonus credits
EnterpriseCustom pricing

Technical Specs

Inputs

Text Prompt, Audio File, Document Image, PDF, URL, Voice Input

Outputs

Transcribed Text, Translated Text, Synthesized Audio, OCR Output, LLM Response, Language Detection Result, Conversational Agent Response

AI Type

Multimodal

Model Architecture

MoE Transformer

Daily Prompts

N/A

Context Length

128K

Output Quality

Accuracy

83%

Content

81%

Reasoning

76%

Company Profile

Company

Sarvam AI (Axonwise Private Limited)

Founded

2023

HQ

Bengaluru, Karnataka, India

Employees

114

Total Raised / Total Funding

$53.8M

Revenue

$3.5M

Valuation

$200M

ARR

$4.2M

CEO

Pratyush Kumar

Overview

Estimated Paid Users

N/A

Current estimate

Total Earnings Till Date

$3.5M

+13.91% from last month

Market Share

0.2%

Current share

Average Session

12

Per active user

Hallucination Rate

22%

Model quality signal

Growth Rate

+20.69%

Monthly active users

Burn Rate

$500K

Total expenses / years active

Paid User Gain

+63.64%

Monthly paid user trend

Profit Analysis

-$13.6M

Total Loss

$38.8M

Total Profit

$0

Performance Metrics

Accuracy

83%

Context

81%

Reasoning

76%

Safety

84%

Benchmarks

Custom Internal Benchmark

84.3%

Custom Internal Benchmark

93.3%

Sarvam AI Models

Sarvam-105B (Indus)

Type: Reasoning

Description: 105-billion parameter MoE LLM trained from scratch on Indic-language-rich datasets. Activates ~9B parameters per token. Supports advanced reasoning, multilingual tasks, mathematics, and coding. Powers the Indus consumer chat app. Open-sourced on Hugging Face under Apache 2.0.

Context Length: 128K tokens

Architecture: MoE Transformer

Sarvam-30B

Type: Text

Description: 30-billion parameter MoE LLM trained from scratch. Activates ~1B parameters per token. Designed for real-time interactive applications. Open-sourced on Hugging Face under Apache 2.0.

Context Length: 128K tokens

Architecture: MoE Transformer

Sarvam Vision

Type: Multimodal

Description: Vision-language model for OCR and document understanding in Indian scripts. Achieved 84.3% on olmOCR-Bench, outperforming Gemini 3 Pro and DeepSeek OCR v2. Supports scanned government documents, historical records, and mixed-script images.

Context Length: 128K tokens

Architecture: MoE Transformer

Saaras V3

Type: Audio

Description: Multilingual speech-to-text model supporting 22 Indian languages with real-time streaming and speaker diarization capabilities. Powers Sarvam's STT API used by enterprise voice applications.

Context Length: 128K tokens

Architecture: MoE Transformer

Bulbul V3

Type: Audio

Description: Text-to-speech model designed for natural prosody across Indian accents, dialects, and scripts. Produces low-latency audio output suitable for IVR systems, voice assistants, and accessibility applications.

Context Length: 128K tokens

Architecture: MoE Transformer

Funding Rounds & Investors

Total Funding

$53.8M

Rounds

3

Series A

$10M

Aug 2025

Follow-on Series A round. Total raised to date reaches $53.8M. Source: Tracxn, Aug 2025.

Series A

$41M

Dec 2023

Led by Lightspeed; largest Series A for an Indian AI startup at the time. Source: Sarvam AI blog announcement, Dec 7 2023.

Seed

$2.8M

Aug 2023

Initial seed round; first institutional capital. Source: Tracxn, Aug 2023.

Founders/Team

PK

Pratyush Kumar

Co-Founder & CEO

VR

Vivek Raghavan

Co-Founder

Direct competitors

No direct competitors available.

Change Log / Major Updates

2026 · Feb 18

At the India AI Impact Summit in New Delhi, Sarvam AI unveils and open-sources Sarvam-30B and Sarvam-105B — both trained from scratch on Indic datasets using MoE architecture. Weights released on Hugging Face under Apache 2.0.

2026 · Feb 20

Sarvam AI launches the Indus consumer chat app on iOS, Android, and web — powered by Sarvam-105B. The app supports voice and text in 22 Indian languages and achieves 50,000+ Play Store downloads within weeks.

2026 · Mar 1

Sarvam launches a structured startup program offering 6–12 months of API credits, priority engineering support, and co-branding opportunities to early-stage Indian AI startups.

Compliance, Integrations & Support

Industry: Not specified

Compliances: Not specified

Integrations: Microsoft Azure, Hugging Face, AIKosh, AI4Bharat, UIDAI Aadhaar, Nokia HMD, Yotta Data Services, E2E Networks, Tata Communications, WhatsApp, REST API, Python SDK, Node.js SDK

Support:email, help center, community forum, enterprise support, discord

Target audience: Enterprise Developers, Government Agencies, Startups, Indian Language App Builders, Voice AI Builders, Fintech Companies, Healthtech Firms, EdTech Platforms, API Developers, Public Sector

Supported languages: Hindi, Tamil, Telugu, Kannada, Malayalam, Bengali, Marathi, Gujarati, Punjabi, Odia, Assamese, Urdu, English, Hinglish

Sarvam AI Acquisitions

PA

Pipable AI

January 1, 2024

N/A

AD

Reviews & Rating

0 reviews

No reviews yet

Be the first to share how Sarvam AI performs for your workflow.

0.0

Accuracy

0.0

Ease of Use

0.0

Output Quality

0.0

Security

0.0

More About Sarvam AI

In August 2023, two researchers who had spent years making AI work for India — not merely work in India — stepped away from academia and government advisory roles to start a company. Pratyush Kumar and Vivek Raghavan had seen the same pattern repeat: India's 1.4 billion people were being served by models trained overwhelmingly on English data, built by companies on the other side of the world, running on infrastructure India did not own.

Building the Full Stack from Scratch

Sarvam AI's ambition is not to wrap an existing open-source model and localise it — it is to own the entire generative AI stack, from foundational model training to inference infrastructure to user-facing interfaces, within Indian borders. The company's February 2026 landmark was the open-sourcing of Sarvam-30B and Sarvam-105B, both trained from scratch using Indian-language-rich datasets and a mixture-of-experts (MoE) architecture. The 105B model, branded Indus in its consumer app form, activated approximately 9 billion parameters per token and supported a 128,000-token context window — trained entirely on government-allocated H100 GPUs inside India.

  • Saaras V3 — multilingual speech-to-text supporting 22 Indian languages with real-time streaming
  • Bulbul V3 — text-to-speech with natural prosody across Indian accents and scripts
  • Sarvam Vision — a vision-language model for OCR and document understanding in Indian scripts, scoring 84.3% on olmOCR-Bench, outperforming Gemini 3 Pro and DeepSeek OCR v2
  • Sarvam Arya — an enterprise AI reliability layer for business-critical deployments

Sovereign AI and the IndiaAI Mission

In April 2025, the Indian Ministry of Electronics and Information Technology selected Sarvam — chosen from 67 applicants — to build India's first indigenous sovereign LLM under the ₹10,372 crore IndiaAI Mission. The government provided subsidised access to 4,096 NVIDIA H100 GPUs, making it possible to train at a scale otherwise out of reach for a startup of Sarvam's size. Partnerships with UIDAI (Aadhaar), Microsoft, and Nokia followed. The Indus consumer app launched in beta in February 2026 with 50,000+ Play Store downloads within weeks, signalling consumer appetite for a home-built Indian AI assistant.

Sarvam AI FAQ's

What is Sarvam AI and what does it do?

Sarvam AI is a Bengaluru-based AI company that builds full-stack generative AI for India. Its products include large language models (Sarvam-30B and Sarvam-105B), speech-to-text (Saaras V3), text-to-speech (Bulbul V3), a vision-language model for document OCR (Sarvam Vision), and the Indus consumer chat app — all optimised for 22 Indian languages.

Is Sarvam AI free to use?

The Indus consumer app is free to use in beta. The developer API provides free credits on signup for experimentation. Beyond the free credits, usage is billed on a pay-as-you-go basis — for example, STT costs approximately $0.35 per hour of audio and TTS costs ₹30 per 10,000 characters. Enterprise plans with custom pricing are also available.

Which Indian languages does Sarvam AI support?

Sarvam AI supports 22 Indian languages including Hindi, Tamil, Telugu, Kannada, Malayalam, Bengali, Marathi, Gujarati, Punjabi, Odia, Assamese, Urdu, and English. The models are also trained to handle code-mixed speech (e.g. Hinglish), which is common in everyday Indian communication.

Has Sarvam AI been selected by the Indian government?

Yes. In April 2025, the Ministry of Electronics and Information Technology selected Sarvam AI — from 67 applicants — to build India's first indigenous sovereign Large Language Model under the IndiaAI Mission. The government provided access to 4,096 NVIDIA H100 GPUs at Yotta Data Services for the model training effort.

How does Sarvam Vision compare to global models on OCR benchmarks?

Sarvam Vision scored 84.3% on the olmOCR-Bench, outperforming Gemini 3 Pro and DeepSeek OCR v2. It also achieved 93.28% on OmniDocBench v1.5. These results reflect the model's specialisation in reading Indian scripts, government documents, and mixed-language printed text.