Sarvam AI Plans Trillion-Parameter Foundation Model

Sarvam's existing flagship is Sarvam 105B, a mixture-of-experts model trained from scratch. Current company documentation lists 105 billion-plus total parameters, a 128,000-token context window and 12 trillion pretraining tokens. The company describes it as a reasoning and agentic model; those specifications and benchmark claims apply to the released 105B system, not to the proposed trillion-parameter model.

The Times of India reported that Sarvam's platform had processed 325 million minutes of voice calls and that one current model scored within four to five percentage points of a much larger model on an unspecified benchmark. Because the report does not identify the benchmark or comparison model, the claim is not enough to establish frontier-level performance.

The Times of India reported that Sarvam also announced an India-hosted inference platform, a Python training SDK, AI-powered Kaze glasses, the Bulbul V4 text-to-speech model, Sarvam Vision 2.0 and Indus, an agentic platform spanning work, voice and coding. The official Epoch page confirms the July 30 and July 31 conference schedule and a session titled "New from Sarvam: Models & Products," but it does not document the trillion-parameter roadmap itself.

Sarvam's current official pricing page lists separate input, cached-input and output rates for Sarvam 105B in Indian rupees. That makes direct "times cheaper" comparisons sensitive to workload mix, currency conversion and competitor tier, so the comparison is not used here as a verified property of the future model.

For practitioners, the consequential questions are training resources, active versus total parameters, multilingual evaluation, inference cost, safety testing and release access. Parameter count alone does not establish capability, and the proposed model cannot be assessed against released systems until Sarvam publishes technical evidence or an independently testable model.