Bengaluru-based Sarvam AI released Sarvam Prism around 3 November 2026, a System-1 decision model on Sarvam 105B that returns calibrated option probabilities in one forward pass. The company reported a median 44 ms on one NVIDIA B200 GPU and an average score of 71.8 across three decision indices.

Weights are open under Apache 2.0 on Hugging Face and via the Sarvam API. Self-hosting needs about 120 GB of GPU memory in FP8. The primary company write-up is on the Sarvam Prism blog. The launch follows earlier Sarvam coverage on Elets CIO, including Vivek Raghavan’s warning on India becoming a digital colony without sovereign AI.

How Sarvam Prism scores options in one pass

Prism is a LoRA fine-tune of Sarvam 105B aimed at discrete choices—routing, classification and agent evaluation—where a calibrated probability per option matters more than generated prose. Sarvam reported that with fixed option ordering, none of 2,000 answers changed when options were shuffled, a stability check for production routing.

On company figures, Prism’s average of 71.8 across three indices was the highest among 25 decision models that finished under 100 ms in their test. On an internal Indian-language decision benchmark it scored 82.5, versus 75.0 for Perplexity Decider v1.2 and 70.8 for Jev 1.14—again company-reported, not independently audited here.

Calibration, languages and independent checks

Sarvam said that at confidence of 90% or higher, actual error was 2.2% in English and 4.4% in Indian languages, below the rate implied by the stated confidence. Inc42 separately cited an Atlan Decision Bench result of 91.5% with a 65 ms median; treat that as a third-party snapshot, not a full audit of every claim.

The release sits alongside India’s wider model push, including IndiaAI Mission foundation models from Soket and ZenteiQ, and enterprise adoption findings such as the TCS Enterprise AI study on how few firms let AI agents act alone.

Where enterprises can apply Prism

Sarvam positions Prism for customer-support routing, document intelligence, security and prompt-injection checks, and AI-agent trace evaluation. On an internal cybersecurity benchmark of 5,517 questions, the company reported 89% accuracy. That security angle lands next to policy work such as CERT-In’s AI security audit rules with 6-hour reporting.

For CIOs, the open Apache 2.0 weights and single-GPU latency pitch matter as much as the headline scores: teams can self-host if they have ~120 GB FP8 capacity, or call the API, while keeping every benchmark labelled as company-reported until broader public evals catch up.

Key takeaways

• Product: Sarvam Prism is a System-1 decision model on Sarvam 105B that returns calibrated option probabilities in one pass, not chat text.
• Speed and score (company): Median 44 ms on one B200; average 71.8 across three indices; Indian-language internal score 82.5.
• Access: Apache 2.0 weights on Hugging Face plus Sarvam API; ~120 GB GPU memory (FP8) for self-host.

FAQ

What is Sarvam Prism?

Prism is Sarvam AI’s System-1 decision model, a LoRA fine-tune of Sarvam 105B that scores a fixed list of options with calibrated probabilities in a single forward pass.

How fast is Prism and how was it scored?

Sarvam reported a median 44 ms on one NVIDIA B200 GPU and an average of 71.8 across three decision indices, the highest among 25 models under 100 ms in its test. Independent figures such as Atlan Decision Bench (91.5% / 65 ms median, per Inc42) should be read separately.

Is Prism open source?

Yes for weights: Apache 2.0 on Hugging Face at sarvam/prism, with API access as well. Self-hosting needs roughly 120 GB of GPU memory in FP8.

Featured image: Press art of a glass prism through a sandstone arch for the Sarvam Prism launch. Photo: The Hindu BusinessLine / Sarvam AI

Be a part of Elets Collaborative Initiatives. Join Us for Upcoming Events and explore business opportunities. Like us on Facebook , connect with us on LinkedIn and follow us on Twitter.

"Exciting news! Elets technomedia is now on WhatsApp Channels Subscribe today by clicking the link and stay updated with the latest insights!" Click here!