---
title: "Sarvam Prism Open-Weight Decision Model Hits 44 ms"
description: "The Bengaluru AI lab released open weights and a hosted API for Prism, which returns calibrated probabilities for fixed choices instead of generating text..."
url: "https://bharat.digital/originals/sarvam-prism-open-weight-decision-model-44ms/index.md"
canonical: "https://bharat.digital/originals/sarvam-prism-open-weight-decision-model-44ms/"
published: "2026-11-03T22:42:07+04:00"
updated: "2026-11-03T22:42:07+04:00"
section: "Originals"
type: "startup"
author: "bharat.digital desk"
image: "https://bharat.digital/img/originals/sarvam-prism-open-weight-decision-model-44ms-auto-51bfe453-og.jpg"
language: "en-IN"
site: "bharat.digital"
---

# Sarvam AI opens Prism, a 44 ms decision model built on Sarvam 105B

*The Bengaluru AI lab released open weights and a hosted API for Prism, which returns calibrated probabilities for fixed choices instead of generating text, aimed at routing, moderation and agent checks.*

By bharat.digital desk. Startup, Startups & Fintech. Published 2026-11-03T22:42:07+04:00. 1 min read.

![Sarvam AI opens Prism, a 44 ms decision model built on Sarvam 105B](https://bharat.digital/img/originals/sarvam-prism-open-weight-decision-model-44ms-auto-51bfe453-hero.webp)

## At a glance

- Company: Sarvam AI
- What it does: Open-weight System-1 decision model and hosted Decisions API built on Sarvam 105B for fast structured choices
- Based in: Bengaluru
- Latest: Launched Prism open-weight decision model
- Date: 2026-11-03

Bengaluru-based Sarvam AI has released Sarvam Prism, a System-1 decision model built on its Sarvam 105B mixture-of-experts backbone. AlphaSignal and AngelNews reported the launch on November 3, 2026. Instead of writing sentences, Prism takes a context plus typed questions with fixed options and returns a calibrated probability for every option in one forward pass. The company positions it for routing, triage, moderation, judging agent runs and security checks such as prompt-injection screening.

AlphaSignal reported a vendor-measured median latency of 44 milliseconds and a 95th percentile of 420 milliseconds on a single NVIDIA B200, using 1,000 sequential requests from a public Decision Index benchmark. Sarvam said Prism averaged 71.8 across three decision indices, which it frames as the highest among models that respond in under 100 milliseconds. AngelNews cited component scores of 67.0 on the Jev Decision Index, 65.9 on a recreation of the Artificial Analysis Decision Index and 82.5 on an Indian-language benchmark. Benchmark figures are company-reported.

Prism is available as open weights on Hugging Face under the Apache 2.0 licence (repository sarvamai/prism) and through Sarvam's hosted Indus Decisions API. Sarvam's documentation lists API pricing at Rs 4 per 1 million input tokens and Rs 0 for output tokens, with support for English, 10 Indian languages and Hinglish. Local inference needs one NVIDIA GPU with FP8 support and about 120 GB of free memory; AlphaSignal noted that Prism's custom FP8 layout is not compatible with AutoModelForCausalLM or general-purpose inference servers, so teams must use Sarvam's stack or adapt their own.

Sarvam said Prism already powers internal workloads such as trace evaluation, prompt guardrails and routing inside its token factory. For answers given with at least 90 percent confidence, the company reported actual error rates of 2.2 percent in English and 4.4 percent in Indian languages, below the error rates its confidence scores would imply. That underconfidence claim is also vendor-reported and matters for automated fraud or security filters where false certainty is costly.

## Sources

1. [AlphaSignal: Sarvam Prism turns a 105B open model into a 44 ms decision engine](https://alphasignal.ai/news/sarvam-s-prism-turns-a-105b-open-model-into-a-44ms-decision-engine)
2. [AngelNews: Sarvam AI launches open-weight Prism model](https://angelnews.in/ai-news/sarvam-ai-prism-105b-decision-model-launch/)
3. [Hugging Face: sarvamai/prism model card](https://huggingface.co/sarvamai/prism)
4. [Sarvam docs: Sarvam Prism model page](https://docs.sarvam.ai/api/token-factory/models/sarvam-prism)

First spotted in a post on X: https://x.com/i/trending/2117634517875691673

Written by the bharat.digital desk from the sources above. Figures are as reported by those sources.