Overview
- Sarvam announced at its Epoch 2026 developer conference on Thursday that it is building a foundation model with more than one trillion parameters and launching an India-hosted inference service to serve enterprise and government customers.
- The startup said it currently has access to about 2,000 NVIDIA Blackwell GPUs and plans to scale to roughly 10,000 GPUs to support model training and domestic inference infrastructure.
- Company benchmarks presented at the event claim Sarvam 105B and new products are far cheaper than rival offerings — for example Sarvam 105B at $0.80 per million blended tokens and Sarvam Code solving tasks for about $2 each — but those performance and price comparisons are reported by Sarvam and lack independent verification.
- New product launches include Vision 2.0 and Vision Edge for OCR and on-device document intelligence (being trialed with Odisha land records), Saras V4 for Indic speech, Bulbul V4 text-to-speech, and Sarvam Code, which is entering beta under the Indus platform.
- Sarvam also named Devendra Singh Chaplot a part-time advisor, opened a San Francisco research office, and framed the push as a move to offer a sovereign, end-to-end AI stack for India while acknowledging significant execution, data-residency and monetisation challenges ahead.