---
title: "Sarvam Prism Open Decision Model Launch | bharat.digital"
description: "The Bengaluru lab says Prism scores 71.8 with 44 ms median latency on a B200 and ships Apache 2.0 weights for enterprise workflows."
url: "https://bharat.digital/originals/sarvam-prism-open-weight-decision-model-105b/index.md"
canonical: "https://bharat.digital/originals/sarvam-prism-open-weight-decision-model-105b/"
published: "2026-11-05T16:50:24+04:00"
updated: "2026-11-05T16:50:24+04:00"
section: "Originals"
type: "startup"
author: "bharat.digital desk"
image: "https://bharat.digital/img/originals/sarvam-prism-open-weight-decision-model-105b-auto-76aba4da-og.jpg"
language: "en-IN"
site: "bharat.digital"
---

# Sarvam AI launches Prism, an open-weight decision model on Sarvam 105B

*The Bengaluru lab says Prism scores 71.8 with 44 ms median latency on a B200 and ships Apache 2.0 weights for enterprise workflows.*

By bharat.digital desk. Startup, Startups & Fintech. Published 2026-11-05T16:50:24+04:00. 1 min read.

![Sarvam AI launches Prism, an open-weight decision model on Sarvam 105B](https://bharat.digital/img/originals/sarvam-prism-open-weight-decision-model-105b-auto-76aba4da-hero.webp)

## At a glance

- Company: Sarvam AI
- What it does: Indian foundation-model lab; Prism is a fast open-weight decision model for enterprise workflows
- Based in: Bengaluru
- Latest: Launched Sarvam Prism open-weight decision model on Sarvam 105B
- Date: 2026-11-03

Bengaluru based Sarvam AI has released Sarvam Prism, a decision model built on its Sarvam 105B foundation model, the company said in a November 3, 2026 blog post, with Economic Times and Inc42 covering the launch. Decision models take a state, a question and a fixed set of options and return a probability for each choice in one forward pass, instead of generating free-form text. Sarvam said Prism is a LoRA fine-tune of Sarvam 105B with a small scoring head, aimed at enterprise tasks such as support routing, document checks, security screening and evaluating AI agent actions.

On company-run tests against 25 decision models, Prism scored an average 71.8 across three indices with a median response time of 44 milliseconds on an Nvidia B200 GPU, the highest average among models under 100 milliseconds in that test set, according to Sarvam's blog and Inc42. Sarvam also said Prism's answers were stable across option orderings on 2,000 questions and that observed error rates stayed inside its stated confidence bounds, including on Indian language tests. Inc42 notes these remain company-reported results pending independent production benchmarks.

Sarvam is releasing Prism weights on Hugging Face under the Apache 2.0 licence and offering the model through its API, the blog and Economic Times reported. The company said Prism runs in FP8 on a single GPU with at least 120 GB of memory, so enterprises can fine-tune it on their own data and host it themselves. Inc42 places the launch alongside other Indian decision-model efforts such as ConvAI Innovations' Laya and PurpleTalk's Sieve, arguing specialised choosers may be a sharper Indian bet than racing only on general chat models.

What could still go wrong: the headline scores come from Sarvam's own suite, and real workflows need calibration on messy private data, GPU cost and failure modes when the option set is incomplete. Open weights help adoption, but buyers will still demand third-party evals before they trust Prism to route money, tickets or security decisions at scale.

## Sources

1. [Sarvam AI: Introducing Sarvam Prism](https://www.sarvam.ai/blogs/sarvam-prism)
2. [Inc42: Sarvam Prism enters India's AI decision model race](https://inc42.com/features/sarvams-prism-enters-indias-growing-ai-decision-model-race/)
3. [Economic Times: Sarvam AI releases Prism decision model](https://economictimes.indiatimes.com/tech/artificial-intelligence/sarvam-ai-releases-prism-a-fast-decision-model-for-enterprises-heres-how-it-works/articleshow/135528004.cms)

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