Sarvam Releases Prism, an Open-Weight Decision Model Built on Sarvam 105B

On November 3, 2026, Indian lab Sarvam released Prism, a rather unusual model: it writes nothing at all. Provide it with a situation (a customer ticket, an agent trace, an incoming message), one or more questions, and a list of possible answers, and it returns a probability for each option in a single computational pass. The weights are available on Hugging Face under an Apache 2.0 license, and the model is also offered via Sarvam’s API.

A Model That Chooses Rather Than Writes

Sarvam categorizes Prism as a “System 1” decision model, referencing the fast, intuitive thinking described by Daniel Kahneman. According to the developer’s blog post, this category was initiated last month by Typeface with its Jev model. The concept mirrors the Audio Decision API introduced by Hume AI on November 2: rather than a free-text response that requires downstream parsing, the model delivers numbers against which software can set hard thresholds.

Technically, Prism is a LoRA fine-tune of Sarvam 105B, the company’s large mixture-of-experts model. A small classification head outputs probabilities instead of generating tokens one by one. Sarvam also altered the attention mask: each option is read independently of the others, ensuring isolated judgments. The model supports three question formats: multi-choice options, yes/no, and rubric scales.

Developer-Reported Performance

Sarvam claims Prism scores an average of 71.8 across three benchmarks with a median response time of 44 milliseconds on a B200 GPU—the highest score among decision models responding in under 100 ms out of the 25 it evaluated. A few models achieve higher scores, but run two to six times slower according to the same comparison chart. These tests were run by Sarvam itself: one of the three benchmarks is an internal test in Indian languages, while another is an in-house recreation of Artificial Analysis’s decision benchmark.

The developer places particular emphasis on calibration: in English, on responses given with at least 90% confidence, Prism predicted an error rate of 3.3% and incurred 2.2%. In Indian languages, competing models tested reportedly made 2.7 to 4.8 times more errors than their reported confidence implied. Sarvam also notes that outputs remain invariant when shuffling option order, and that Prism outperforms Claude Haiku 5.5 when evaluating agent traces (78.5% accuracy versus 70.7%). None of these benchmarks have been independently reproduced yet.