Indian artificial intelligence startup Sarvam has announced plans to build a model containing more than one trillion parameters as it expands its domestic AI infrastructure and model-development capabilities.
The announcement was made during Epoch 2026, Sarvam’s artificial intelligence conference in Bengaluru. The proposed model will represent a significant increase in scale from the company’s existing 30-billion and 105-billion-parameter models.
Sarvam did not announce the commercial release of the trillion-parameter model. The company is building the system as part of its broader effort to develop frontier-scale artificial intelligence models in India. Further details about its architecture, number of active parameters and expected availability have not yet been disclosed.
Apart from its language-model roadmap, Sarvam announced the commercial availability of Sarvam Vision Edge, a document-intelligence platform designed to process documents locally. The platform can extract and organise information from PDFs, scanned records and other documents while preserving elements such as tables, layouts and reading order.
Edge-based processing allows data to be handled on local devices or systems instead of being sent continuously to remote cloud infrastructure. This approach can reduce latency and support organisations handling confidential or sensitive documents.
Sarvam’s vision technology supports English and India’s 22 scheduled languages. Its applications include document digitisation, optical character recognition, table conversion and the processing of government, financial and enterprise records.
The company also said its AI models are priced 5.5 times lower than comparable offerings from global competitors. This pricing comparison is based on Sarvam’s assessment and has not been independently verified. The cost difference may also depend on the model, workload, token consumption and deployment method being compared.
Sarvam is building a full-stack AI platform covering language models, speech recognition, speech synthesis, translation, document intelligence and on-device processing. Its existing model portfolio includes Sarvam 30B and Sarvam 105B, along with speech and vision systems developed for Indian languages.
Founded by Vivek Raghavan and Pratyush Kumar, Bengaluru-based Sarvam develops artificial intelligence models and applications focused on Indian languages and use cases. The company is also participating in India’s sovereign AI programme.
