Sarvam releases Prism, an open-weight decision model with company-reported 44 ms benchmark latency
Sarvam’s open-weight Prism scores fixed options rather than generating text. The company reports 71.8 average accuracy and 44 ms median latency on a B200 in its own tests; calibration and production performance require validation on each deployment’s data.
Most attention in AI goes to models that write, talk and reason at length. Much enterprise automation needs something narrower: a fast, reliable answer to a closed question. Should this insurance claim be escalated? Which of five categories does this support ticket belong to? Is this transaction suspicious? On 3 November 2026, Bengaluru-based Sarvam AI released Sarvam Prism, a model built specifically for this kind of decision-making.
What Prism is
Sarvam calls Prism a "System One" decision model, borrowing the psychologist's term for fast, intuitive thinking as opposed to slow, deliberate reasoning. Technically it is a LoRA fine-tune of Sarvam 105B, the company's flagship open-weight large language model.
Its design is what sets it apart. A typical chatbot generates an answer one token at a time. Prism instead receives a state, a question and a fixed set of options, and returns a probability for each option in a single forward pass, without generating any text. Because the model only scores options it has been given, it cannot produce a free-form answer outside the allowed choices, a type of error that worries risk teams. It also makes Prism much faster and gives it a more predictable output format, which matters when a model sits inside a business workflow and not in a chat window.
Speed and benchmarks
Sarvam reports that Prism:
- has a median response time of 44 milliseconds on a B200 GPU in Sarvam’s benchmark setup;
- scores an average of 71.8 across three benchmark indices, including Sarvam's recreation of the Artificial Analysis Decision Index and an internal Indian-language benchmark;
- achieves, in Sarvam's words, the highest average score under 100 milliseconds among the 25 models Sarvam tested.
These are Sarvam-run measurements across the tested indices, not independent evidence that every production request will take 44 ms. Fixed-option outputs can still select the wrong answer.
Calibrated confidence
The feature that may matter most to enterprises is calibration. Sarvam says Prism is trained to produce confidence scores that track how likely it is to be right. A business can then set a threshold: decisions above it are automated, and those below it go to a human reviewer. Any company deploying AI in regulated sectors such as banking, insurance or healthcare needs this kind of control. A model that is confidently wrong is a liability. A model that knows when it is unsure can be governed.
Open weights and modest hardware
Prism follows Sarvam's open approach. Reports say the weights are on Hugging Face under the Apache 2.0 licence, and the model is also available through Sarvam's hosted API. Sarvam says it runs in FP8 precision on a single GPU with at least 120 GB of memory. That is still serious hardware, but it is within reach of Indian enterprises and government departments that want to run AI on their own infrastructure and not send data abroad.
The Sarvam 105B foundation
The primary announcement identifies Sarvam 105B-10B as the base, with a LoRA fine-tune and a scoring head that returns option probabilities in one forward pass.
Since then, Sarvam has been building a product line on that base. It includes speech models such as Bulbul, enterprise tools, and now Prism. The company has also signed government and defence partnerships . Prism shows a clear strategy: rather than competing only on general-purpose chat, Sarvam is turning its sovereign base model into specialised products that companies can pay for.
Why it matters for India
There are three reasons the release is significant for India's AI ecosystem:
- Indian-language decisions. Many Indian business processes run in Hindi, Tamil, Bengali and other languages. A decision model tested on an Indian-language benchmark targets a need that global providers often treat as secondary.
- Cost and latency. At about 44 ms per decision, automation that would be too slow or costly with a text-generating LLM becomes practical: real-time fraud checks, call-centre routing and form triage.
- Sovereign deployment. Open weights let banks, public-sector bodies and defence users run the model entirely on their own hardware. This is central to the argument for sovereign AI. Critics have asked whether releasing weights publicly undermines "sovereignty". Supporters reply that it is what makes deployment under domestic control possible.
What to watch
Independent evaluation is the next test. If third-party benchmarks confirm Sarvam's speed and accuracy, Prism could become a common building block for Indian enterprise automation. Adoption will also depend on how easily companies can fine-tune it on their own decision data, and on whether the calibration holds outside Sarvam's test sets. With other IndiaAI-backed teams also releasing models, Indian AI is moving from showing it can build large models to showing that those models are useful in production.
Sources
- https://www.sarvam.ai/blogs/sarvam-prism
- https://x.com/SarvamAI/status/2117621415520448838
- https://analyticsindiamag.com/ai-news/sarvam-ai-launches-prism-a-jev-like-decision-model-built-on-sarvam-105b
- https://www.unite.ai/sarvam-ai-releases-open-weight-prism-decision-model/
- https://explainx.ai/blog/sarvam-prism-open-weights-decision-model-105b-44ms-calibrated-confidence-apache-2-2026
- https://www.newsbytesapp.com/news/science/sarvam-ai-launches-prism-enterprise-ai-for-automating-company-decisions/tldr
- https://artificialanalysis.ai/articles/sarvam-105b-Sarvam-30b-everything-you-need-to-know
- https://en.wikipedia.org/wiki/Sarvam_AI