Overview
- Sarvam announced Thursday that it is building a trillion‑plus parameter foundational AI model in India but did not provide a timeline or technical details for training, architecture, or datasets.
- The company launched Sarvam Inference, an India‑hosted platform that runs Sarvam’s 105B model and select open models on domestic infrastructure to meet enterprise and government data‑residency needs.
- Sarvam presented company benchmarks and pricing that claim its 105B model outperforms rivals on voice and agent tasks and costs roughly $0.80 per one million blended tokens, a figure the company says is far cheaper than competing offerings; these claims await independent verification.
- New commercial multimodal products include Vision 2.0 for OCR and document extraction, Vision Edge for on‑device document intelligence used in an Odisha land‑records research pilot, Saras V4 for expanded Indic speech recognition, and Bulbul V4 for more expressive text‑to‑speech.
- Sarvam signalled international expansion with a planned San Francisco office, named Devendra Singh Chaplot as an advisor, and said a recent $234 million fundraising round led by HCL Tech pushed its market value to about $1.5 billion, positioning the startup to pursue enterprise and public‑sector contracts.