Every enterprise runs on decisions — thousands of small, fast judgements a day: route this ticket here, flag that document, check this agent’s work, block that suspicious prompt. Sarvam AI, the Bengaluru company building India’s own AI stack, has just released a model built for exactly that moment: Sarvam Prism, a decision model that answers in 44 milliseconds, and whose weights are open for any enterprise to customise and run on its own infrastructure.
Prism is what Sarvam calls a System One model, after the fast, intuitive mode of judgement that handles most of what people decide in a day. Instead of writing out sentences, it does something much narrower and much faster: given a situation, a question and a set of options, it returns a probability for each option in a single forward pass. No token-by-token generation, no waiting for a paragraph — just a calibrated answer, ready for a workflow to act on.
Fast — and accurate at that speed
The headline numbers come from Sarvam’s own testing, run on the same hardware for every model it compared. Across three decision indices, the company says Prism scored an average of 71.8 while responding in a median 44 milliseconds on a single NVIDIA B200 GPU — the highest score among the 25 decision models it tested that answer in under 100 milliseconds. A few models scored higher overall, the company says, but they took between two and six times as long to respond.
That speed did not come free — it was engineered. Prism started life at 664 milliseconds per request, running in 16-bit precision across two B200 GPUs. By converting the weights to FP8 so the model fits on a single GPU, running all 128 experts in one grouped kernel, capturing the forward pass as a CUDA graph and folding the FP8 conversion into the expert kernels, Sarvam brought the median request down to 44 milliseconds — 15 times faster, with no drop in accuracy, the company says. Prism runs in FP8 on one GPU with at least 120 GB of memory, and handles up to 131,072 tokens per question, so long traces can be evaluated in one pass.
Confidence an enterprise can act on
Speed alone is not enough to automate a decision — a business needs to trust the model’s confidence. That is where Prism’s most distinctive result lies. The company says that across every confidence threshold it tested, from 50% to 95%, in English and in Indian languages, Prism made fewer errors than its confidence scores predicted. On answers given with at least 90% confidence, it made errors in 2.2% of English responses against a predicted 3.3%, and 4.4% of Indian-language responses against a predicted 5.4%.
On Sarvam’s internal Indian-language benchmark, Prism scored 82.5, ahead of Perplexity Decider v1.2 at 75.0 and Jev 1.14 at 70.8, according to the company’s tests. And it is steady in another way enterprises care about: when the order of the options was shuffled, Prism’s answers stayed the same across all 2,000 questions tested once the options were placed in a fixed order before scoring — so the same question does not get a different answer just because the options arrived in a different order.
More Read
Built for real enterprise work
Prism is already at work inside Sarvam itself, the company says — judging agent runs, guarding prompts, routing decisions in its token factory and testing speech-to-speech systems. On recorded agent runs, it scored 78.5% across all questions, ahead of Claude Haiku 5.5 at 70.7%, and scored a full agent run in about half a second on a single GPU. On an internal cybersecurity benchmark of 5,517 questions, covering phishing, jailbreak attempts and attack classification, the company says Prism scored 89.0%, ahead of Clef at 84.5% and Jev 1.14 at 78.6%.
The early applications Sarvam names read like a tour of the modern enterprise: customer support routing, document intelligence, security checks and evaluation of AI-agent performance, plus speech-to-speech testing, prompt-injection guardrails, intent routing and model routing. Anywhere a workflow needs a fast, structured judgement — not an essay — Prism is designed to slot in.
Open weights, made in India
Perhaps the most important part of the release is what Sarvam is giving away. Prism’s weights are released under the Apache 2.0 licence on Hugging Face, and the model is also available through Sarvam’s API. Every enterprise decides differently, the company says — so it is releasing Prism as an open, calibrated base that enterprises can train on their own data and outcomes, and run on their own infrastructure.
Prism is a LoRA fine-tune of Sarvam 105B, the company’s large Mixture-of-Experts model, trained with two equally weighted objectives that make confident mistakes costly — one reason, the company says, for its careful calibration. It is the latest step from a Bengaluru team that has moved from foundation models to consumer hardware with the Kaze smart glasses, and now to the quiet, fast machinery that lets enterprises put AI to work in their own workflows. Forty-four milliseconds at a time, India is building the decision layer of the AI age — and this one is open for anyone to build on.
Keep reading: Pay Only for the Tokens You Use: Inside E2E Networks’ Blaze, India’s Pay-Per-Token AI Inference Service · Maths, Code and 22 Languages: Inside Soket AI’s EKA-1, India’s New Homegrown Foundation Model · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
