Sarvam AI Launches Prism: A 44-Millisecond Decision Model with Open Weights
Sarvam AI just dropped Prism, an open-weights decision model that routes tasks and analyzes documents in 44 milliseconds.
Models · Source: The Hindu BusinessLine
What happened
Indian AI startup Sarvam AI launched Prism. It is a decision-making model built on their Sarvam 105B base model. Unlike chat models that write essays, Prism handles fast and structured decisions. It picks from fixed options for tasks like customer support routing and document analysis.
Speed is the main feature. Prism delivers responses in 44 milliseconds when running on an NVIDIA B200 GPU. Sarvam tested 25 decision models that respond in under 100 milliseconds. Prism took the top spot with an average score of 71.8 across three benchmarks.
The model is highly accurate in both English and Indian languages. On Sarvam's internal Indian-language benchmark, Prism scored 82.5. It beat Perplexity Decider v1.2 and Jev 1.14. It also provides reliable confidence estimates. When Prism is 90 percent confident, it makes errors in just 2.2 percent of English responses.
Key facts
- 44 milliseconds — Response time on an NVIDIA B200 GPU
- 71.8 — Average score across three benchmarks
- 82.5 — Score on Sarvam's internal Indian-language benchmark
- Apache 2.0 — The open-source license used for Prism's weights
- 89 percent — Accuracy on an internal cybersecurity benchmark
Why it matters
Builders can now run high-speed routing and evaluation locally. Chat models are too slow and unpredictable for split-second enterprise workflows. Prism solves this by focusing entirely on structured choices. You get consistent answers every time, even when the order of options changes.
Data privacy becomes a default rather than a luxury. Sarvam released Prism under the Apache 2.0 license. Enterprises can train the model on their own data and deploy it on their own infrastructure. Companies no longer have to send sensitive customer data to third-party APIs just to route a support ticket.
For builders
Run local AI agent evaluations
Prism is built to evaluate AI agent performance. You can run it on your own servers to grade your agents without paying API fees or leaking proprietary data.
Automate enterprise security checks
The model scored 89 percent accuracy on an internal cybersecurity benchmark of 5,517 questions. Security teams can deploy this locally to flag threats in milliseconds without exposing internal logs.
Build fast customer support routers
You can replace slow LLM calls in your support pipeline. Prism routes tickets in 44 milliseconds, saving compute costs and reducing wait times for end users.
My take
AI is not just about generating text anymore. It is about making split-second decisions where data privacy is non-negotiable. Sarvam giving away the weights under Apache 2.0 means you get speed and total control over your data. This is how enterprise AI actually scales.
Original reporting: The Hindu BusinessLine. This is my rewrite and opinion.