Sarvam announced at its Epoch 2026 conference that it is building a base model exceeding one trillion parameters and aims to compete with leading global labs across domains from code to cybersecurity. The Indian startup used its keynote to showcase several product updates — from the Sarvam 105B model to new vision and voice tools — and to outline a strategy focused on domestic hosting and local use cases.
A “trillion-plus” model developed in India
In the opening remarks, co-founder Pratyush Kumar said, “we are building a model with more than one trillion parameters right here in India,” adding that the goal is to be competitive in code, cybersecurity, simulation, and science. To put that scale in context, the press release cites Kimi K3 as an example of a current large open-weight model (2.8 trillion parameters). “Parameters” refer to the model’s weights that determine its behavior and its ability to generalize.
Sarvam 105B: claimed performance at lower cost
Sarvam also announced improvements to its 105B model, particularly for conversational agents and voice calling. Benchmarks provided by the company claim that Sarvam 105B outperforms GPT-5.4 Mini and Gemini 3.5 Flash on tasks such as instruction following and voice capabilities. On pricing, Sarvam states a blended cost of $0.80 per million tokens — reportedly 5.5× cheaper than GPT-5.4 Mini at $4.50 per million tokens.
(Note: the source text ended mid-sentence; this translation reflects the available content.)
