Sarvam Prism: open AI decision model answers in 44 milliseconds

Sarvam AI, the Indian start-up that builds its own AI models, released Sarvam Prism on Tuesday, 3 November 2026. The company describes it as a "System One" decision model: an AI system meant to make quick, narrow choices inside business software, such as which team a customer complaint should go to or whether a message should be blocked, rather than write long answers. Built on Sarvam's own 105-billion-parameter model, Sarvam 105B, Prism scores 71.8 on average across the tests Sarvam ran, with a median response time of 44 milliseconds on one Nvidia B200 graphics processor, according to the company. Sarvam says the model's weights are open for enterprises to download and run on their own computers, and that it is also available through Sarvam's paid API.

Most people meet AI as a chatbot that writes paragraphs. Inside companies, a lot of AI work is smaller and more frequent: deciding which department a ticket belongs to, whether a transaction looks risky, or which of several AI models should handle a request. These are yes-or-no or pick-one-of-five decisions, made millions of times a day. Speed matters because every delay adds up across a workflow, and cost matters because a large chatbot model is expensive to call for a one-word answer. Prism is Sarvam's attempt to sell a specialised tool for exactly that job.

The second selling point is calibration. Sarvam says Prism gives a confidence score with each decision, and that the score is honest: among English answers given with at least 90% confidence, the model predicted a 3.3% error rate and recorded 2.2%, according to the company. A well-calibrated confidence score lets a business decide when to trust the machine and when to send a case to a human.

For India, the release matters because Sarvam is one of the few Indian companies training large models of its own. An open-weight model that runs on a company's own hardware can appeal to banks, insurers and government departments that do not want customer data to leave their premises. Sarvam says the model can run in FP8, a compact number format, on a single GPU with at least 120 GB of memory.

All the scores are Sarvam's own. No independent body has yet tested Prism, and secondary coverage notes that a slower model, Mammoth-1, scored higher at 75.2 but took 284 milliseconds. Prism's lead holds only among models that answer in under a tenth of a second.

Peepals calculation: on Sarvam's averages, Prism's 71.8 is 3.8 points above Perplexity Decider's 68.0, about 5.6% higher (3.8 ÷ 68.0 ≈ 0.056), and 10.7 points above Microsoft-Decision-1's 61.1, about 17.5% higher (10.7 ÷ 61.1 ≈ 0.175). On calibration, the recorded error rate of 2.2% was a third lower than the 3.3% the model predicted (1.1 ÷ 3.3 ≈ 0.33). Our calculation from the reported figures.

Short-term: Developers and enterprise IT teams can try Prism through Sarvam's API or by downloading the weights. Early users will test whether the company's speed and accuracy figures hold on their own data.

Long-term: If independent tests confirm the claims, fast and cheap decision models could take over routine choices inside customer support, fraud checks and content moderation, with large chatbot models used only where long answers are needed. That would also give Indian companies a home-grown option in a field where US firms such as Microsoft and Perplexity offer similar products.

Who is affected: Indian enterprises running customer-service, security and AI-agent systems; developers building on open models; and rival AI providers in the decision-model space.

Questions and Answers