Sarvam says it is making 1 trillion parameter AI model, seeks to challenge OpenAI and Anthropic

Published 30 July 2026 · tech

AI startup Sarvam has made a series of announcements during its AI conference, called Epoch 2026. At the event, the Indian AI lab claimed that

AI startup Sarvam has made a series of announcements during its AI conference, called Epoch 2026. At the event, the Indian AI lab claimed that its models can deliver strong performance at a fraction of the cost of global frontier models. From Sarvam 105B to the launch of Sarvam Vision 2.0 and Bulbul V4, here is everything you need to know. Read Full Story During the keynote, Sarvam co-founder Pratyush Kumar said that the startup was building a foundational AI model with over a trillion parameters in India "We are very happy to announce that we are building a trillion-plus parameter model right here in India. We are building them from scratch to be competitive in coding, cybersecurity, simulation, science and more," he said. Parameters refer to the dataset used to train AI. The more the parameters the stronger the AI can perform. For context, Kimi K3, the largest open-weight model right now, is a 2.8 trillion parameter model. That is, with a trillion parameter model, Sarvam may get closer to leading AI labs like Anthropic and OpenAI.

Sarvam 105B is 5.5 times cheaper than competition Sarvam has made improvements to its flagship 105B model, when it comes to how the model works with AI agents. According to benchmarks shared by the company, Sarvam 105B is better at voice call capabilities, instruction following, and other tasks when compared to OpenAI’s GPT-5.4 Mini and Google’s Gemini 3.5 Flash. AI costs have quickly become a concern for companies, and Sarvam says that the 105B is 5.5 times cheaper than GPT-5.4 Mini. Sarvam 105B costs $0.80 per one million blended tokens, compared with $4.50 for GPT-5.4 Mini. On the other hand, the Sarvam model is 11 times cheaper than Google’s Gemini 3.5 Flash which costs $9 for a million blended tokens. Saravm says that the model is being hosted in India, and is available for voice products. The company said, "If you are building a voice product today, Sarvam is the cheapest, fastest and most scalable solution," and added that competing voice offerings from ChatGPT are not currently available while Gemini is not available at comparable scale.

Sarvam Vision 2.0 and Vision Edge Sarvam also announced Vision 2.0, the next-generation of Sarvam Vision. Vision 2.0 focuses on optical character recognition (OCR) capabilities for enterprise document processing. That is, the AI model can read and interpret documents and text. Vision 2.0 brings major improvements when compared to the original Vision 1, which had managed to outperform models from OpenAI and Google in OCR. Vision 2.0 is also better at recognising Indian handwritten text and extracting tables or key values. Sarvam also announced commercial availability of Sarvam Vision Edge, its document intelligence platform that can run locally on any device. Vision Edge is being used by the Odisha government on a research initiative to digitise and extract information from land records. Sarvam Bulbul 4 and Saras v4 For speech recognition, Sarvam has launched Saras V4 that gets even better at recognising Indic languages. Apart from languages like Hindi and Bengali, Saras V4 can also transcribe languages that other AI models cannot, such as Odia, Sanskrit and Manipuri.