Sarvam AI Launches Open-Weight Prism Model With 44ms Response Time and 89% Cybersecurity Benchmark Accuracy

Sarvam AI has introduced Sarvam Prism, a new System One decision-making model built on its Sarvam 105B foundation model. The model is designed to support fast and reliable decision-making across enterprise workflows, including AI agent evaluation, customer support automation, cybersecurity guardrails and document intelligence.

According to the company, Sarvam Prism achieved an average score of 71.8 across three decision-making benchmarks, with a median response time of 44 milliseconds on a single NVIDIA B200 GPU. The company said this was the highest average score among the 25 decision models evaluated that delivered responses in under 100 milliseconds.

Sarvam Prism Records 71.8 Average Score Across Three Benchmarks

Sarvam AI evaluated Prism against 24 other decision models using three benchmarks: the Jev Decision Index, an internally recreated Artificial Analysis Decision Index and an internal Indian-language benchmark.

Prism recorded scores of 67.0 on the Jev Decision Index, 65.9 on the Artificial Analysis benchmark and 82.5 on the Indian-language benchmark, resulting in an average score of 71.8.

The company said the model outperformed several competing models in the combined assessment while maintaining a median response time of 44 milliseconds. Models such as mammoth-1 and GLiDE 1.2 achieved higher overall average scores of 75.2 and 73.2, respectively, but recorded considerably longer response times of 284 milliseconds and 116 milliseconds.

Sarvam Prism also demonstrated competitive performance in Indian languages, recording a score of 82.5, compared with 75.0 for Perplexity Decider v1.2 and 70.8 for Jev 1.14.

Focus on Confidence and Decision Reliability

Sarvam AI said Prism is designed to provide reliable confidence estimates, an important requirement for enterprises looking to automate decisions without continuous human intervention.

In tests involving answers with at least 90% confidence, Prism recorded an actual error rate of 2.2% against an expected error rate of 3.3% in English. For Indian languages, its actual error rate was 4.4%, compared with an expected rate of 5.4%.

The company said Prism’s actual error rates remained within its predicted confidence levels across the confidence thresholds tested, ranging from 50% to 95%.

This approach is intended to help businesses determine when an AI-generated decision can be automated and when a particular case requires human review.

Consistent Decisions Regardless of Option Order

Sarvam AI also evaluated whether changing the order of available options would influence the model’s decisions.

Across 2,000 questions, Prism maintained consistent answers when the options were presented in a fixed order. The model recorded no changes across four repeated submissions of the same requests.

When options were shuffled, the model’s answers changed in 4% of cases when processed in their original presentation order. Sarvam addressed this by arranging options in a fixed order before scoring them and returning the corresponding probabilities under the original option names. With this approach, the model recorded no changes across the 2,000 questions tested.

Applications in AI Agents and Cybersecurity

Sarvam Prism has also been evaluated for enterprise applications involving AI agent monitoring, security and workflow automation.

In tests covering 700 recorded AI agent runs and 1,770 evaluation questions, Prism achieved an overall accuracy of 78.5%. It recorded 75% accuracy in identifying whether an agent successfully completed a task, 86.4% in determining whether human review was required and 69.6% in identifying which agent caused a failure.

The model also achieved 89% accuracy in Sarvam AI’s internal cybersecurity benchmark, which included 5,517 questions across nine tasks. The evaluation covered identifying phishing-related content, detecting jailbreak attempts and classifying security attacks using MITRE frameworks.

Other potential applications include speech-to-speech systems, prompt injection detection, customer support intent routing, model routing and document intelligence.

Built on Sarvam 105B With Open Weights

Sarvam Prism is developed through a Low-Rank Adaptation (LoRA) fine-tuning process using the Sarvam 105B model. Instead of generating responses token by token, it processes the provided context, question and available options in a single forward pass.

A scoring component then assigns probabilities to the available options. Its training process uses cross-entropy loss with 5% label smoothing and Brier loss to improve probability calibration and reduce confident prediction errors.

Sarvam AI said several optimisation techniques helped reduce the model’s median response time from 664 milliseconds to 44 milliseconds, representing approximately a 15-fold improvement without a reported reduction in accuracy.

These optimisations include converting model weights to FP8 precision, using grouped processing for all 128 experts, CUDA Graph execution and merging the fine-tuning adapter into the model weights.

The company is releasing Sarvam Prism’s weights on Hugging Face under the Apache 2.0 licence, allowing enterprises to fine-tune the model using their own datasets and deploy it on their infrastructure. The model runs in FP8 precision on a single GPU with at least 120 GB of memory and is also accessible through Sarvam’s Indus API.

With Sarvam Prism, the company is targeting enterprise applications where decision accuracy, response speed and confidence calibration are important for deploying AI-driven workflows at scale.