Sarvam AI releases Prism, a 105B-based decision model that claims 44 ms responses and a 71.8 average score

Sarvam AI has launched Prism, a System 1 decision model built on its Sarvam 105B model, which the company says returns responses in 44 milliseconds and averages 71.8 across three benchmarks. The Bengaluru-based startup released the weights on Hugging Face under the Apache 2.0 licence. Prism is also available through the Sarvam API (pricing not yet disclosed).

How Prism works

Prism is a LoRA fine-tune of Sarvam 105B. Conventional language models generate responses token by token. Prism instead processes the input in a single forward pass and returns a probability score for each available option.

Sarvam says the model is aimed at enterprise workflows built from sequences of decisions: intent routing, customer support, model selection and document intelligence. The Apache 2.0 licence lets enterprises fine-tune it on their own data and deploy it on their own infrastructure.

Deployment requirements

• Precision and hardware: FP8 on a single GPU with at least 120 GB of memory
• Input length: up to 1,31,072 tokens per question, as Sarvam states it
• Agent trace evaluation: about half a second for an entire trace on a single GPU
• Access: open weights on Hugging Face, plus the Sarvam API

Benchmark results (company-reported)

Sarvam tested Prism against 24 other decision models on a single NVIDIA B200 GPU. The company says Prism recorded the highest average score among models with response times below 100 milliseconds. In a blog post, Sarvam said Prism "sets a new point on the Pareto frontier of accuracy and speed among the 25 decision models we tested."

Benchmark	Prism score
Jev Decision Index (Sarvam's recreation of the Artificial Analysis Decision Index)	67
Artificial Analysis benchmark	65.9
Internal Indian-language benchmark	82.5
Average across the three	71.8

These results come from Sarvam's own testing, including an internal benchmark and an internal recreation of a third-party index.

Calibrated confidence

Sarvam says Prism is trained to produce calibrated confidence scores, so enterprises can decide when an output is reliable enough to automate and when a human should step in. When Prism was at least 90% confident, its error rate was 2.2% on English questions and 4.4% on Indian-language questions. The company says Prism "makes fewer errors than its confidence scores predict, across every threshold we tested from 50% to 95%, in English and Indian languages."

Sarvam also tested sensitivity to answer ordering. Across 2,000 questions, Prism's answers stayed the same when options were shuffled, after the model arranged them in a fixed order before scoring.

Agent and security evaluations

• Identifying whether an agent run required human review: 86.4% accuracy
• Determining whether a task had been completed successfully: 75% accuracy
• Internal cybersecurity benchmark of 5,517 questions across nine tasks: 89% accuracy. The tasks covered phishing detection, jailbreak attempts and classification of attacks using MITRE ATT&CK frameworks.

What this means

Prism is built for choosing among fixed options rather than generating text, and Sarvam positions it for steps in a pipeline where latency and reliability matter. The calibrated confidence scores are the central design feature: they give teams a threshold for routing low-confidence cases to humans. The reported error rates at 90% confidence are Sarvam's own figures, and the Jev Decision Index and Indian-language benchmark are Sarvam-built, so independent evaluation would be needed to verify the accuracy-speed claims. The open Apache 2.0 weights and the single-GPU FP8 footprint mean teams can run and fine-tune Prism in their own environments and test it on their own data.