Sarvam is building 1T+ parameter AI model in India
What's the story
Indian AI start-up Sarvam has announced that it is building a one trillion-plus parameter foundational model in the country. The move is aimed at developing frontier AI capabilities locally. The company also revealed a pricing strategy that it claims is 5.5 times cheaper than leading global models. Pratyush Kumar, the co-founder of Sarvam, made the announcements at an event in Bengaluru.
Competitive edge
Competing with global giants
The new model from Sarvam is being developed from scratch in India and is aimed at competing in coding, cybersecurity, scientific research, and simulation workloads.
This comes as AI companies are racing to build larger foundation models while also lowering inference costs to attract enterprise customers.
The company's flagship Sarvam 105B model costs $0.80 per one million blended tokens, which is significantly cheaper than OpenAI's GPT-5.4 Mini ($4.50) and Gemini 3.5 Flash ($9).
Performance comparison
Performance on par with larger models
Sarvam also claims its 105-billion parameter model performs on par with much larger models at a fraction of the cost.
The company compared Sarvam 105B against GLM 5.2 (753B) across multiple enterprise benchmarks, including tool use, robustness, consistency, and process execution.
Despite the difference in model size, the performance gap remained narrow.
Product launch
Document intelligence platform launched
Sarvam also announced the commercial launch of Sarvam Vision Edge, its document intelligence platform.
The company has partnered with the Odisha government on a research project where the vision model is being used to digitize and extract information from land records.
Sarvam Vision 2.0 was also introduced. It improves optical character recognition (OCR) capabilities for enterprise document processing.