Sarvam AI
India's sovereign LLM, built for 22 Indian languages
Overview
Sarvam AI is the Bengaluru startup the IndiaAI Mission selected in April 2025 to build India's sovereign large-language-model ecosystem, now one of twelve organisations it supports. Its Sarvam-30B and Sarvam-105B models are open under Apache-2.0 and power the Indus assistant. A $234M first close of a $300M Series B at a $1.5B valuation (June 2026) made it an AI unicorn.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 27 Sep 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with Sarvam AI
- A multilingual customer-support chatbot that handles Hindi, Tamil, Telugu, and Marathi queries — including code-mixed Hinglish inputs — without routing to a human agent
- Automated transcription and translation of audio calls from any of 22 Indian regional languages into structured text for CRM logging
- A natural-sounding Hindi or Tamil voice IVR flow using Bulbul V3 TTS for banking or government helplines
- Extraction of tables, invoices, and handwritten forms written in Indian scripts from scanned PDFs into structured JSON or Markdown via the Vision OCR API
- An AI tutor that explains school-curriculum maths problems in Hinglish, drawing on Sarvam-M's large gains (~+21% on math, including romanized GSM-8K) over its base model
- Integration of a government-scheme FAQ assistant — citizens ask questions in their native language (e.g., Kannada or Odia) and receive accurate scheme-eligibility answers
- A regional-language content pipeline that drafts, translates, and publishes product descriptions or news summaries across 10+ Indian languages at near-English token cost and speed
ASEAN Perspective
Sarvam AI in Southeast Asia
A flagship APAC sovereign-AI project — most relevant for Indian-language workloads and South Asian markets, and a reference point for ASEAN governments weighing their own sovereign-LLM strategies. Open Apache-2.0 weights make it freely self-hostable.
Sarvam is the most credible vehicle for India's sovereign-AI push: Sarvam-30B and Sarvam-105B were trained end-to-end on IndiaAI compute, cover 22 Indian languages, and ship under Apache-2.0, so builders can self-host without licence anxiety. Government backing plus a June 2026 $234M round at a $1.5B valuation — HCLTech led — buys it real runway, and the Indus consumer app gives the models a public face. It isn't chasing GPT or Gemini on frontier capability, the 2025 Sarvam-M episode left trust scars the from-scratch 2026 models are still repairing, and low-resource Indic languages lag the big ones. For India-first products, Indic-language interfaces or data-residency mandates, it's the obvious first call; outside those use cases it's a niche pick.
What we found
February 20, 2026 was the day Sarvam had to prove the sceptics wrong, and it largely did: Sarvam-105B — a mixture-of-experts model activating 10.3B parameters per token, with a 128K context window — launched in beta as the Indus app on iOS, Android and the web, and both it and the smaller Sarvam-30B were open-sourced under Apache-2.0 on Hugging Face in early March. Crucially, pre-training, fine-tuning and reinforcement learning all ran in India on GPU capacity provisioned through the IndiaAI Mission. That answers the loudest criticism from 2025, when the Sarvam-M release was widely panned as a Mistral fine-tune dressed up as sovereign AI, and alleged bot-inflated Hugging Face download numbers dented community trust.
The money followed. In June 2026 Sarvam announced $234M as the first close of a $300M Series B at a $1.5B post-money valuation, with HCLTech committing $150M as lead strategic investor, making it India's newest AI unicorn — and one of twelve organisations the IndiaAI Mission now supports, rather than the only one.
Capability claims still deserve the self-reported asterisk: the roughly 90% pairwise win rates come from Sarvam's own evaluations — against open models such as Qwen3-Next, GLM-4.5-Air and GPT-OSS-120B, with Gemini 3 as the judge rather than an opponent — and practitioners report weaker quality on low-resource languages than on Hindi or Tamil. There is no meaningful G2 or Capterra footprint — this is a developer-and-government story, not a SaaS one. Indian developers read as cautiously optimistic: nobody expects it to outreason the frontier labs, but for Indic-language products, speech (Saaras STT, Bulbul TTS), Indic OCR and India data-residency requirements, there is no more credible home-grown option.
API pricing is now published, in rupees: the 105B model costs ₹29.28 per million input tokens and ₹73.20 per million output, speech-to-text ₹30 an hour and text-to-speech ₹3 per thousand characters. Updated 105B variants for conversation and agent work shipped in August 2026, through the API only.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- Chosen by India's IndiaAI Mission to build the national sovereign LLM
- Sarvam-105B is India's first fully home-trained open-source LLM (105B params)
- Supports 22 Indian languages
- Became a unicorn in June 2026 with a $234M round led by HCLTech ($1.5B valuation)
Frequently asked questions
About this listing
This entry was compiled from publicly available data including Sarvam AI's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with Sarvam AI unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to Sarvam AI directly →
Spotted something out of date? Suggest an update →
Related reading
Alternatives to Sarvam AI
More in LLMs & Chat
