Sarvam releases Prism, an open-weight decision model that answers in 44 milliseconds

Sarvam AI has added a new category to its model line-up with Sarvam Prism, a model designed not to converse but to decide. Built on the Sarvam 105B base, Prism is tuned for fast, structured judgements: which queue a support ticket belongs in, whether a document passes a compliance check, whether an AI agent's output is acceptable. Sarvam reports an average response time of 44 milliseconds on a single Nvidia B200 GPU.

The company's own benchmarking places Prism at an average of 71.8 across three evaluation sets, which it says is the highest among 25 decision models it tested that respond in under 100 milliseconds. On an Indian-language benchmark it scores 82.5, against 75 for Perplexity Decider v1.2 and 70.8 for Jev 1.14. On an internal cybersecurity set of 5,517 questions it reached 89 per cent accuracy, compared with 84.5 per cent for Clef and 78.6 per cent for Jev 1.14.

Sarvam has put particular emphasis on calibration. For answers given with at least 90 per cent confidence, the measured error rate was 2.2 per cent in English and 4.4 per cent in Indian languages, both below what the model itself predicted. The company says this relationship held from a 50 per cent to a 95 per cent confidence threshold, and that answers remained unchanged across all 2,000 test questions when options were presented in a fixed order, a check against position bias.

The weights are released under the Apache 2.0 licence, so enterprises can train Prism on their own data and run it inside their own infrastructure. It is also available through Sarvam's API. The figures are company-reported via a syndicated ANI release and have not yet been independently reproduced, which buyers should bear in mind.

Why it matters

Most production AI spend is not on long conversations but on millions of small classification and routing decisions. A calibrated, Indic-capable decision model that returns in under 50 milliseconds and can be self-hosted addresses a gap that general chat models fill expensively and slowly.

Indicops view

Prism is the sort of component we look for when designing grievance triage, document intake and agent guardrails for Indian departments: narrow, fast, auditable and deployable on-premise. The calibration claims are the most important feature if they hold, because a model that knows when it is unsure can route uncertain cases to a human rather than guess. We will test Prism on live routing workloads against our current classifiers and report what we find.