by Sarvam AIIN
Give India capable AI models it controls itself, trained on its own compute and good at its own languages, that anyone can download and run.
Open weights since March 2026; 105B runs the API and Indus (30B retired); a 1T+ model is next.
Updated 9 Mar 2026Checked 11 Oct0 updates this week
The 105B model has 128 expert sub-networks; a router picks 8 per token, so about 10B weights do the work and serving stays cheap.
The 105B model uses latent attention, which shrinks what it must remember about earlier text, so 128K-token inputs fit in less GPU memory.
Training leaned on the 10 most-spoken Indian languages, in native script, in Latin letters and mixed with English, the way people type.
Pre-training, supervised fine-tuning and reinforcement learning all ran on 4,096 H100s hosted by Yotta under the IndiaAI Mission.
| Spec | Sarvam 30B / 105B |
|---|---|
| Total parameters (largest) | 105BR (reported) |
| Active per token (105B) | 10.3BR (reported) |
| Active per token (30B) | ~2.4BR (reported) |
| Experts (105B) | 128R (reported) |
| Context window (105B) | 128K tokensR (reported) |
| Pre-training data (105B) | 12T tokensR (reported) |
| Training GPUs | 4,096 NVIDIA H100R (reported) |
| Licence | Apache 2.0R (reported) |
R reported by the company
Sarvam AI
Indian-language LLMs, voice AI
Sarvam AI is a Bengaluru lab building AI models from scratch for Indian languages, from large language models to speech and document reading. The Indian government picked it in 2025 to build the country's first sovereign model, and in June 2026 it became a unicorn.
