Sarvam Prism: India’s 44-Millisecond Decision AI Explained

Sarvam AI released Sarvam Prism on 3 November 2026, and its headline number is deliberately narrow: a median response of 44 milliseconds on a single NVIDIA B200 graphics processor, with an average score of 71.8 across three decision benchmarks. Prism does not write essays, hold conversations or generate images. It takes a situation, a question and a set of options, and returns a probability for each option in a single forward pass. That is the whole point — and the reason it should not be sold as a chatbot rival.

Quick answer: Sarvam Prism scores 71.8 in 44 milliseconds by deciding, not chatting. What the open-weight model does, where it beats rivals, and what the numbers do not prove.

What is Sarvam Prism?

Prism is what Sarvam calls a System One decision model, borrowing Daniel Kahneman’s term for fast, intuitive judgement. Much enterprise work, Sarvam argues, is a sequence of such decisions: route this ticket, flag this document, check this agent run, approve this step. A model built for those calls does not need to generate sentences; it needs to choose, quickly, and to say how confident it is.

It is a LoRA fine-tune of Sarvam 105B, the company’s 105-billion-parameter mixture-of-experts base model, specifically the Sarvam 105B-10B base. Sarvam also released Prism’s weights under the Apache 2.0 licence, so an enterprise can train the calibrated base on its own data and run it on its own infrastructure, or use Sarvam’s API instead. Weights are published on Hugging Face. Prism runs in FP8 precision on one GPU with at least 120 GB of memory, according to Sarvam’s launch post.

That open-weights line is the story’s second half. A decision model locked behind one vendor’s API can be fast and still be a dependency. An open-weight model can be audited, fine-tuned and hosted where the data already lives — provided the enterprise has the hardware and the discipline to test it on its own workloads first.

How Prism decides without generating text

A conventional language model answers by generating tokens one after another, which is why even a short reply takes time and why the same question can drift into different wording. Prism skips that step. It reads the state, the question and the options in one forward pass, and a small scoring head returns a probability for each option. No token-by-token generation happens at all.

Two design choices matter for anyone planning to trust those probabilities. First, the state and question form one block that is read in both directions, while each option sees that block and itself but never the other options, and every option starts from the same position. Sarvam says isolating the options makes each judgement independent, and that the bidirectional reading adds about 3 points on the Jev Decision Index. Second, the training objective combines cross-entropy with 5% label smoothing — which discourages complete certainty — and a Brier loss, which makes confident mistakes costly. Sarvam attributes Prism’s calibration to those two choices.

Speed comes from engineering, not just from skipping text generation. Prism started at 664 milliseconds per median request in 16-bit precision across two B200 GPUs. Converting the weights to FP8, running all 128 experts in one grouped kernel, capturing the forward pass as a CUDA graph, folding the FP8 conversion into the expert kernels and merging the fine-tuning adapter into the model weights brought the median to 44 milliseconds — 15 times faster, with no drop in accuracy, Sarvam says. That is a vendor-reported optimisation story, not an independent reproduction, and it should be read that way.

The numbers, benchmark by benchmark

Sarvam tested Prism and 24 other decision models itself, on the same benchmarks, timing every model it could run on the same B200 hardware using each developer’s own inference code. Two models available only through an API, Jev 1.14 and Microsoft-Decision-1, are plotted at their publicly reported times. The headline 71.8 is the average across three indices:

IndexPrismPerplexity Decider v1.2Jev 1.14ClefMicrosoft-Decision-1
Jev Decision Index (37 public benchmarks)67.067.760.761.757.8
Artificial Analysis recreation (Sarvam’s version, with a curated hard-question set including an “I don’t know” option)65.961.261.954.855.6
Internal Indian-language benchmark82.575.070.871.270.0
Average71.868.064.562.661.1

Response times, as medians per request on one B200 GPU, were 44 ms for Prism, 117 ms for Perplexity Decider v1.2, 222 ms for Clef, and reported times of 196 ms for Jev 1.14 and 85 ms for Microsoft-Decision-1. Sarvam’s claim, stated precisely, is that Prism has the highest average score among decision models responding in under 100 milliseconds across the 25 models tested. Sarvam itself adds that a few models score higher overall but take between two and six times as long to respond — so “fastest good decision” is the honest headline, not “best model, full stop”.

The latency figure also needs its test conditions attached. It is a median over 1,000 requests sent one at a time, not under concurrent load. Independent coverage of the launch notes a 95th-percentile latency of 420 ms on the same hardware. A business running Prism at scale, with longer inputs and many simultaneous requests, should expect different numbers and should measure them before promising a 44-millisecond experience to users.

Why calibration matters more than speed

Speed is easy to market; calibration is what lets a business automate safely. An enterprise automates a decision by acting on it only above a confidence threshold, and that is safe only if the confidence output is honest — if a model that claims 90% confidence is actually wrong more than 10% of the time, the threshold is a trap.

On that test, Sarvam reports Prism as underconfident across every threshold tested from 50% to 95%, in English and Indian languages. For English answers given with at least 90% confidence, Prism predicted 3.3% errors and made 2.2%; for Indian-language answers at the same threshold, it predicted 5.4% errors and made 4.4%. Its rivals, by contrast, underestimated their errors: Perplexity Decider, Microsoft-Decision-1 and Jev 1.14 made 1.5, 1.8 and 2.7 times as many English errors as their confidence scores predicted, and 2.7 to 4.8 times as many in Indian languages.

Prism also proved unusually stable when the options were shuffled. Across 2,000 questions, Sarvam says Prism — after putting options into one fixed order before scoring and returning each probability under its original name — changed 0 answers, while Jev 1.14 changed 12.3% when shuffled and Perplexity Decider changed 10.3%. Order-sensitivity sounds academic until a support-routing system treats the same ticket differently because the queue listed the categories in a different order. It is a genuine robustness result, with the caveat that it comes from Sarvam’s own test set.

What Prism is already being used to judge

Sarvam lists early real-world uses that fit the decision-model shape: a speech-to-speech harness, trace evaluations, prompt-injection guardrails, intent routing for customer support, model routing and document intelligence. Two of its own evaluations show why those fits are plausible.

On agent trace evaluation — judging 700 recorded agent runs, or 1,770 questions, on whether the task succeeded, whether a human review was needed, and which agent caused a failure — Prism scored 78.5% overall, ahead of Clef and Jev 1.14 at 76.4%, Microsoft-Decision-1 at 75.7% and Claude Haiku 5.5 at 70.7%. A full agent run takes about half a second on a single GPU, Sarvam says, against a median 2.5 seconds per run for Claude Haiku 5.5 through its API.

On an internal cybersecurity benchmark of 5,517 questions across nine tasks, built from vulnerabilities published since July 2026, Prism scored 89.0%, ahead of Clef at 84.5%, Jev 1.14 at 78.6% and Microsoft-Decision-1 at 76.1%. That benchmark is internal and not yet public — Sarvam says it will expand the set and release it publicly in the coming weeks — so treat the 89% as a vendor claim awaiting independent checking, not as a settled security result.

What open weights actually change — and what they do not

Apache 2.0 weights mean an enterprise is not renting a black box by the token. It can fine-tune Prism on its own historical decisions, run it inside its own perimeter, and inspect what it is deploying. For regulated or data-sensitive workflows — a bank routing complaints, a hospital triaging internal requests, a government helpdesk — that control is often worth more than a few extra benchmark points, as recent Apple developer work also shows.

It does not remove the hard parts. Open weights do not supply labelled training data, a threshold policy, a fallback route for low-confidence calls, or an audit trail for decisions that go wrong. A 44-millisecond decision is only useful if the organisation knows what happens at 45 milliseconds when confidence is low: who reviews it, how fast, and what the cost of a wrong auto-decision is. The licence also does not make Prism a general-purpose model. It answers from a fixed set of options; it does not write, reason aloud or handle an open-ended question that has no option list yet. Teams that need free-form answers still need a chat model alongside it. (For a contrasting case of control inside a closed ecosystem, see our Siri AI skills guide linked in Sources.)

Our take: a five-question trial before you trust it

This article was desk-researched from Sarvam’s launch post and same-day reporting; Prism was not hands-on tested for this piece. If you are considering a trial, run it through these five questions first — they are the original checklist this explainer adds, and each one maps to a failure mode the launch data actually shows.

  1. Can you state the decision as options? If the job cannot be written as “given this state and question, choose among these options”, Prism is the wrong tool. Start with routing, triage or review-flagging, not open-ended drafting.
  2. What is your confidence threshold, and who owns the calls below it? Set the threshold before the trial, and route everything below it to a human. Calibration only helps if low-confidence answers actually stop.
  3. Does option order change your current system’s answers? Re-run a sample of past decisions with shuffled option orders. If your current tooling is order-sensitive, Prism’s fixed-order scoring is a real, measurable win — verify it on your own data, not Sarvam’s.
  4. What does latency look like on your hardware, under your load? Reproduce the median and the 95th percentile on your own GPU and request mix. A single-B200 median is a starting point, not a service-level promise.
  5. What will you fine-tune on, and how will you know it helped? Hold out recent decisions, fine-tune on the rest, and compare calibrated accuracy before and after. If fine-tuning does not beat the base model on your held-out set, ship the base model — or do not ship at all.

Worked example: a support team routes tickets into billing, technical, account access or feedback. Prism returns probabilities for the four options; tickets above 90% confidence auto-route, tickets below it go to a human queue with the top two options shown. The team then measures not just routing accuracy but the share of tickets that needed re-routing after auto-assignment — the metric that decides whether the threshold is right. That measurement loop, more than the model itself, is what turns a fast decision into a safe one.

Prism sits in a useful middle that the week’s other benchmark news sharpens. SemiAnalysis’s GDPx, covered in our earlier explainer on the best AI scoring just 12% on real professional work, shows how far frontier agents still are from finishing multi-hour expert tasks unsupervised. Prism is not trying to finish those tasks at all. It is trying to make the small, repeated judgements inside them fast, calibrated and cheap — and to let enterprises own that layer. If that framing holds up under independent testing, it matters more than any single benchmark point. If it does not, the open weights at least let buyers find out for themselves.

Frequently Asked Questions

What is Sarvam Prism?

Sarvam Prism is a System One decision model released by Indian company Sarvam AI on 3 November 2026. It is a LoRA fine-tune of the Sarvam 105B base model that returns a probability for each option in a single forward pass, instead of generating text token by token.

How fast is Sarvam Prism, really?

Sarvam reports a median response of 44 milliseconds per request on one NVIDIA B200 GPU, measured over 1,000 requests sent one at a time. That is a vendor figure on specific hardware under sequential load, not a consumer-device or high-concurrency promise; independent coverage notes a 95th-percentile latency of 420 ms on the same setup.

What does the 71.8 score mean?

It is the average of Prism’s scores across three indices Sarvam used: 67.0 on the Jev Decision Index, 65.9 on Sarvam’s recreation of the Artificial Analysis Decision Index, and 82.5 on an internal Indian-language benchmark. Sarvam claims it is the highest average among the 25 decision models it tested that respond in under 100 milliseconds.

Is Sarvam Prism open source?

Its weights are released under the Apache 2.0 licence and published on Hugging Face, and it is also available through Sarvam’s API. Apache 2.0 allows enterprises to train it on their own data and run it on their own infrastructure, but it does not supply training data, a deployment setup or a support guarantee.

Is Sarvam Prism a ChatGPT rival?

No. Prism does not hold conversations or generate free-form answers. It chooses among a fixed set of options for a given question. It is best understood as a fast decision layer inside larger workflows — routing, triage, evaluation and guardrails — rather than a general chatbot.

Sources and Methodology

Claims in this article rest on Sarvam AI’s own launch post, “Sarvam Prism: a decision model built on Sarvam 105B” (3 November 2026), cross-checked against ANI wire reporting carried by The Tribune (4 November 2026), Analytics India Magazine’s launch coverage (4 November 2026) and The Hindu BusinessLine’s enterprise report (4 November 2026). All benchmark, latency and calibration figures are Sarvam’s own, labelled as vendor-reported where they appear; no independent benchmark reproduction existed at the time of writing, and the article was desk-researched, not hands-on tested. Verification date: 4 November 2026.

Leave a Reply

Your email address will not be published. Required fields are marked *