Bengaluru-based Sarvam AI has announced plans to build a trillion-plus parameter foundation model entirely from scratch in India, positioning it to compete directly with frontier systems from OpenAI, Google,
and Anthropic while undercutting them sharply on price.
The announcement was made at Epoch 2026, Sarvam's first developer conference held in Bengaluru, where the company unveiled its biggest product roadmap to date. Cofounder Pratyush Kumar told the audience at the event that the company was building the trillion-plus parameter model right in India, adding that it was being trained from scratch to be competitive in coding, cybersecurity, simulation, and scientific research. He said the model is expected to arrive within six months, according to the company.
Alongside the announcement of the upcoming frontier model, Sarvam also revealed new pricing for its existing Sarvam 105B model. The upgraded model will now cost $0.80 per million blended tokens, which the company said makes it roughly five and a half times cheaper than OpenAI's GPT-5.4 Mini, priced at $4.50, and more than eleven times cheaper than Google's Gemini 3.5 Flash, priced at 9 dollars per million tokens.
New voice tools take centre stage
Sarvam also used the event to showcase upgrades to its voice technology stack. Bulbul V4, its text to speech model, drew attention during live demos for its ability to produce natural, cinematic sounding speech complete with emotional cues like laughter, excitement, and emphasis, moving away from the flat, robotic tone typical of older systems. Its speech to text counterpart, Saras V4, was also unveiled with multi-speaker support, allowing it to separate and transcribe overlapping conversations, a feature aimed squarely at making meeting recordings and call transcriptions more usable in real-world settings.
Building the full stack in India
Beyond the models themselves, Sarvam announced the launch of an India-hosted inference service, allowing developers to run leading AI models on domestic servers rather than relying on infrastructure based overseas. Cofounder Vivek Raghavan described the move as part of the company's broader push toward what he called token sovereignty, referring to serving a larger share of the AI computing needs generated within India through local infrastructure.
The company also announced the opening of a new office in San Francisco as it looks to expand its global footprint, and confirmed the appointment of Devendra Singh Chaplot, previously associated with Mistral, as an advisor to help guide its frontier model development.
The announcements come at a moment of heightened attention on India's AI ambitions, with the country accounting for a significant share of global usage of tools like ChatGPT and Claude. Sarvam's push to build sovereign, cost-efficient AI infrastructure signals a broader industry shift toward reducing dependence on foreign AI providers, even as the trillion-parameter race among global labs continues to intensify.
