Sarvam AI: India's Own Open-Source AI Model, Explained
Most AI tools people use in India, ChatGPT, Gemini, Claude, are built by American companies. Sarvam AI, a Bengaluru startup, built something different: Sarvam-30B and Sarvam-105B, open-source AI models trained entirely in India and designed around Indian languages from the ground up. We explain what that actually means.
⚡ Quick facts
- Released: unveiled at the India AI Impact Summit in February 2026; weights open-sourced in March under Apache 2.0
- Models: Sarvam-30B (about 2.4B active parameters) and Sarvam-105B (10.3B active), both mixture-of-experts
- Trained on: Domestic compute, backed by the ₹10,372-crore IndiaAI Mission
- Languages: 22 scheduled Indian languages plus English; 128K-token context on the 105B model
See it running
Video: Fahd Mirza, "India Enters the AI Race: Running Sarvam-30B Locally."
📊 Key points from the video (skip the watch)
- A hands-on, technical video that installs and tests Sarvam-30B running locally.
- Also reviews Sarvam-105B, and both are open-source Indian models.
- Aimed more at developers who want to run the models themselves than at a general audience.
Built entirely in India
Sarvam-30B and Sarvam-105B were unveiled at the India AI Impact Summit in New Delhi in February 2026, and Sarvam open-sourced both in March under the Apache 2.0 licence. What sets them apart isn't just where the company is based: the models were trained from scratch rather than built on top of a foreign base model, with the training pipeline (pretraining, fine-tuning and reinforcement learning) run in India on compute provided through the government's ₹10,372-crore IndiaAI Mission.
Why the language support matters
The Sarvam-105B model card claims state-of-the-art performance for its size across 22 Indian languages, including Hindi, Tamil, Telugu, Bengali, Marathi and Kannada, and Storyboard18 reports that a large share of training went into the ten most widely spoken ones. Global models often handle these languages less naturally, and they cost more tokens to process. This is a genuine opening for exactly the kind of regional-language content and tools this audience needs, well beyond what most Western AI products prioritize.
Is this something you can actually use?
If you're a developer, the weights are on Hugging Face and AI Kosh, so you can download and run them yourself or fine-tune them, and both models are also available through Sarvam's API. Sarvam uses them in its own products: Sarvam-30B powers Samvaad, its conversational agent platform, and Sarvam-105B powers the Indus AI assistant.
What "open-source" means here
The models are released as open weights on Hugging Face, which means you can download them, run them on your own hardware and fine-tune them for a specific job. That is different from a hosted service where you only send requests to someone else's servers. For a company handling sensitive data, running the model in its own environment can be the main attraction.
Why "sovereign" matters
Training the whole pipeline on domestic compute means India does not depend on a foreign provider for the model itself. That matters for government and regulated sectors that prefer data and infrastructure to stay in the country. It does not by itself make the model better than the global leaders. Sarvam reports strong scores for the 105B model (90.6 on MMLU, 78.7 on GPQA Diamond, 98.6 on Math500, 45.0 on SWE-Bench Verified), but these are company-run, the model card lists no Indian-language benchmark scores, and we found no independent reproduction, so test it on your own tasks before relying on it.
Our take: what it takes to run it yourself
“Open weights” only helps if you can run them. We estimated the memory each model needs from the parameter counts on the Sarvam-30B and Sarvam-105B model cards (about 2 bytes per parameter at 16-bit precision, about 0.5–0.6 bytes at 4-bit, before memory for the context):
- Sarvam-30B at 16-bit: roughly 60 GB of weights, so one 80 GB data-centre GPU such as an H100 or A100.
- Sarvam-30B at 4-bit: roughly 16–18 GB, within reach of a single 24 GB consumer GPU or a high-memory Mac, if a quantized build is available. With only about 2.4B parameters active per token, it should generate quickly once loaded.
- Sarvam-105B at 16-bit: roughly 210 GB, so a multi-GPU server. Sarvam’s own examples split it across 4 GPUs with SGLang or 8 with vLLM.
Two practical notes from the model cards: SGLang supports the models directly, while native vLLM support was still an open pull request, so vLLM needs Sarvam’s patch or fork. For most teams, the sensible path is to prototype on Sarvam’s API, then self-host Sarvam-30B only if data residency or cost at scale requires it.
Who should try it
- Developers: building chat, search or voice tools for Indian-language users.
- Businesses: that want customer support in several Indian languages, or need to keep data in India.
- General users: can try the Indus assistant instead of setting anything up.
Related reading
Indian-language support is spreading across the industry: see how Google AI Mode now works in Hindi and nine other languages, and how NotebookLM added Hindi and Tamil.
Sources
- Hugging Face: Sarvam-105B model card
- Hugging Face: Sarvam-30B model card
- Storyboard18: Sarvam AI releases India-built 30B and 105B open-source AI models
- YourStory: Sarvam AI unveils two new LLMs
Frequently asked questions
Can I run Sarvam-30B on my own computer?
Possibly. At 16-bit precision it needs roughly 60 GB of GPU memory, but a 4-bit quantized build needs roughly 16–18 GB, which fits a 24 GB consumer GPU or a high-memory Mac. Sarvam-105B needs a multi-GPU server.
What is Sarvam AI?
A Bengaluru startup that built Sarvam-30B and Sarvam-105B, open-source AI models trained entirely on Indian infrastructure.
How is it different from ChatGPT or Gemini?
It is trained with a focus on 22 Indian languages, trained in India, and its weights are open under Apache 2.0, so anyone can download and run it.
Can I use it now?
Yes. Developers can download the weights from Hugging Face or use Sarvam's API; everyday users can try Sarvam's Indus assistant, which runs on Sarvam-105B.
Why does this matter?
It's a step toward India having its own core AI infrastructure and better support for Indian languages.