Sarvam AI, India’s full-stack artificial intelligence (AI) startup announced its working to build an indigenous one trillion plus parameter foundational model that will share capabilities of frontier world AI models in coding, cybersecurity and other domains.
The Bengaluru based startup also launched Sarvam Vision 2.0, a next gen document AI/vision language model customized to read handwritings in Indian languages and Bulbul V4 (Text-to-Speech) model that turns text into natural-sounding speech including emotions such as laughter, along with other major announcements at its flagship conference Sarvam Epoch in Bengaluru on Thursday.
Sarvam also claimed their existing flagship model Sarvam-105B is priced 5.5 times cheaper than leading global models such as GPT-5.4 Mini.
At the inauguration of Sarvam Epoch, Pratyush Kumar, co-founder of Sarvam AI said the foundational model will be built from scratch in India.
“I am happy to announce we are investing in a trillion plus parameters models right here in India building them from scratch to become competent on a range of tasks including coding, cybersecurity, simulations, science etc..sarvam has has access to the largest blackwell cluster in india, we invested and now we are now able to run these clusters right from bare metal,” said Pratyush Kumar.
Sarvam claimed that its flagship AI model Sarvam-105B is now better at handling complex tasks, supports voice, and costs $0.80 per million blended tokens, putting it 5.5× cheaper than GPT-5.4 Mini, while also announcing another 30% price cut.
Sarvam Epoch 2026 – Livestream
Sarvam Epoch is India’s flagship AI conference, bringing together leaders from enterprise, startups, government, research, and technology to understand where AI is headed, what is becoming possible, and what it means for India.…
— Sarvam (@SarvamAI) July 29, 2026
“Sarvam’s goal is to be able to produce the largest share of tokens which India consumes right here in the country and of course make it available to the world. We are announcing an inference platform running right in India which will serve sarvam models but a range of open frontier models, bringing the best of innovation that is out there at globally competitive price points. You shouldn’t have to pay a tax to use models here in India, he said.
We want Indian developers and enterprises to use our models but also use the frontier open models deployed right here in India with data security of the highest order, reliability and price that are globally competitive,” Pratyush mentioned.
Kumar also welcomed Devendra Singh Chaplot, AI Scientist and founding member of Mistral and Thinking Labs, as adviser for Sarvam AI.
Addressing the conference, Chaplot said that there is a lot of misconceptions regarding the compute cost involved in building a frontier AI model from scratch.
“A lot of people think building a frontier model and to avail resources required for it, is impossible to get in India but that’s not true, take the case of Kimi K3, its very much possible in terms of compute to do this in India,” added Chaplot.
At the event, Co-founder of Sarvam- Vivek Raghavan mentioned that Sarvam inference is a platform to serve SOTA open weight models from Indian infrastructure. These models will be Saaras V4, Bulbul V3, Vision, Sarvam Translate, Sarvam dub, GLM 5.2, Gemma 431B, Sarvam 105B.
Also Read: Sarvam Receives $234M in Funding for Frontier AI Research
