Sarvam AI
Sarvam Prism answers business decisions with probabilities, not prose
India’s Sarvam AI has released an open decision model that picks from the options you give it in about 44 milliseconds and reports how confident it is.
Sarvam AI, an Indian company that describes itself as a full-stack sovereign AI platform, released Sarvam Prism on 3 November 2026. Prism is a decision model: instead of writing an answer, it chooses one from a list you supply and tells you how confident it is.
How a decision model works
You give it a context, such as a support ticket, and one or more questions, each with a fixed set of options: which team should handle this, is it urgent, how upset is the customer on a scale. Prism returns a probability for every option in a single pass through the model.
It does not generate text, so the answer is always one of your options.
Prism is built on Sarvam’s 105-billion-parameter model. The category is new: TypeSafe introduced it a month earlier with a model called Jev, named after the fast, intuitive ‘System One’ thinking described by the psychologist Daniel Kahneman. Prism maps options into a fixed order before scoring, and Sarvam reports that shuffling the options changed none of 2,000 test answers.
Speed and calibration
Averaged over three test sets, the Jev Decision Index, a recreation of the Artificial Analysis decision index and an internal Indian-language benchmark, Sarvam reports a score of 71.8 with a median response time of 44 milliseconds on one NVIDIA B200. The two models above it in Sarvam’s table scored 75.2 and 73.2, but took 284 and 116 milliseconds. On Indian languages Prism scored 82.5, the best in Sarvam’s comparison.
Sarvam puts more weight on calibration. On English benchmarks, for answers given with at least 90% confidence, Prism predicted a 3.3% error rate and made 2.2%. A model whose confidence you can trust is one you can set a threshold on.
The weights are released under Apache 2.0. Running them takes one GPU with FP8 support and about 120 GB of free memory, such as a B200 or H200, plus roughly 106 GB of disk space. Sarvam also offers Prism through its hosted API.
What it means for your business
A large share of day-to-day operations is small, repeated decisions: which queue a ticket goes to, whether an invoice matches its purchase order, whether a lead is worth a call. A general chatbot can make those calls, but it is slower, costs more per call and can wander off format. A decision model returns one of your options with a probability attached.
The probability is what makes it work inside a team. You set the bar: confident calls go straight through, uncertain ones land with a person who can see what the model thought. That’s humans and AI working as one team, with the routine handled and the judgement calls kept with people. Sarvam says it released the weights so enterprises can train Prism on their own data and run it on their own infrastructure, which matters to any business that wants its decision data kept onshore.
Sources
- Sarvam Prism (opens in a new tab) · Sarvam AI
- Sarvam-Prism model card (opens in a new tab) · Sarvam AI on Hugging Face
General information only, not legal, financial or professional advice. Check the sources for your circumstances.