Lead Analysis
Sarvam launches Prism: an open-weights decision model built for Indian enterprise
Friday, November 6, 2026: Sarvam AI has released Prism, a decision-making model that answers in 44 milliseconds on an NVIDIA B200, scored highest among the 25 sub-100ms decision models it tested, and ships under Apache 2.0 — deployable entirely on an enterprise’s own infrastructure.
Built on Sarvam’s 105B base model, Prism targets high-volume structured decisions — customer-support routing, document intelligence, security checks and evaluation of AI agents — and posts stronger Ind-language and confidence-calibration scores than the global comparison models in Sarvam’s own tests. It is the first concrete product conversion of Sarvam’s 2026 capital position.
Prism recorded an average score of 71.8 across three benchmarks while responding in 44 milliseconds, which Sarvam says is the highest among the 25 decision models it tested that answer in under 100 milliseconds. On Sarvam’s internal Indian-language benchmark it scored 82.5, against 75 for Perplexity Decider v1.2 and 70.8 for Jev 1.14, and on an internal cybersecurity benchmark of 5,517 questions it scored 89% versus 84.5% and 78.6% for the same pair. These are vendor-run tests, not independent evaluations, and should be read as such.
The confidence story is the part enterprises will actually deploy against. In tests where answers were given with at least 90% confidence, Prism made errors in 2.2% of English responses and 4.4% of Indian-language responses — below its own predicted error rates at every threshold from 50% to 95%. It also returned unchanged answers across all 2,000 questions when option order was shuffled, a stability property that matters for regulated decision logs where a re-run must not flip the outcome.
The Apache 2.0 release is the structural news. Enterprises can fine-tune Prism on their own data and run it on their own infrastructure, with the API as an alternative. For Indian BFSI, government and GCC estates, that is the sovereignty argument in product form: decision flows can run on-premises, with neither data nor decision records leaving the organisation. The carried Tracxn data shows Sarvam captured 87% of 2026 Indic AI capital; Prism is that capital converted into a shippable, deployable line of product.
Prism also creates a new procurement category in India: the sub-100ms decision model, priced and licensed for on-prem deployment, competes directly with the global decision-model wave (Perplexity Decider, Jev) and gives Indian enterprises a homegrown reference point on latency, calibration and Indic-language fit. The agent-evaluation use case is the sleeper — as AI agents proliferate inside Indian IT delivery, the tool that audits other agents becomes a governance purchase, not a convenience.
The launch lands as Indian markets deliver their sixth straight rally, with IT leading the charge after a two-day wobble — Sensex +2,009 to 79,310 and Nifty +576 to 24,537 on November 5, with FIIs returning as net buyers. The AI-services complex that sells agentic decision workflows to global clients now has a homegrown reference model to point at; watch whether Prism-based on-prem deployments start appearing in Q3 earnings calls and BFSI digital-operations tenders in the coming weeks.
