Sarvam AI has announced its biggest expansion yet as the Bengaluru-based startup looks to strengthen India’s position in the global artificial intelligence race. At its Epoch 2026 event, the company said it is building a one-trillion-parameter foundation model from scratch in India while unveiling a new platform that will allow developers and enterprises to build AI models tailored for Indian languages and use cases. The startup also announced aggressive pricing for its existing AI models, claiming they are significantly cheaper than competing global offerings. Adding to the announcements, Sarvam brought onboard AI researcher Devendra Singh Chaplot, who most recently worked at Elon Musk’s xAI, marking a major talent gain for the company.
Speaking at the launch event in Bengaluru, co-founder Pratyush Kumar said Sarvam AI is developing a one-trillion-parameter foundation model in India that will target coding, cybersecurity, scientific research and simulation workloads. He further added that the model is being built from scratch and will be competing with leading global AI systems.
The company also announced pricing for its Sarvam 105B model, stating that the model costs $0.80 per one million blended tokens, compared with $4.50 for OpenAI’s GPT-5.4 Mini and $9 for Gemini 3.5 Flash. According to Sarvam, this makes its offering about 5.5 times cheaper than GPT-5.4 Mini and more than 11 times cheaper than Gemini 3.5 Flash.
The company launched Sarvam Vision Edge, a document intelligence platform, alongside Sarvam Vision 2.0, an improved optical character recognition for enterprise document processing, during the event. Talking about the new tools, Kumar said that the Odisha government is already using the technology to digitise and extract information from land records.
Also read: Asus Pad T3201 with 12.2-inch OLED display launched in India: Price, specifications and availability
Aside from the abovementioned, the company also introduced Epoch Builder Edition. It’s a platform designed to help developers, researchers and enterprises build, fine-tune and deploy AI models for Indian use cases.
The platform includes new multilingual large language models with 7 billion and 70 billion parameters that support more than 10 Indian languages, including Hindi, Kannada, Tamil, Marathi and Bengali. It also provides GPU infrastructure, training tools, curated datasets, deployment options and safety testing features.
Sarvam said the platform is aimed at banks, government departments, startups and research institutions that want to build AI applications using their own data. The company also announced partnerships with three IITs and two state governments for pilot projects in education and public service delivery. Epoch Builder Edition will reportedly enter private preview in August 2026, with a wider rollout planned in the fourth quarter of the year.
Lastly, Sarvam said that the AI researcher Devendra Singh Chaplot has joined the company as a part-time advisor. Chaplot most recently worked at Elon Musk’s xAI and has also been part of the founding teams at Mistral AI and Mira Murati’s Thinking Machines Lab.
Addressing the audience, Chaplot said India needs AI models that can understand the country’s languages and cultural diversity. He added that while AI models may become outdated over time, the ability to build and train them remains valuable. Based in Palo Alto, Chaplot will work with Sarvam’s new foundational AI research lab in San Francisco.