By Laura Anderson, Welp Magazine. Published and last updated November 5, 2026.
Blink and you will miss it. A human blink lasts around 100 to 400 milliseconds. Sarvam Prism, a new AI model from India, answers in 44.
Bengaluru-based Sarvam has launched Prism, and it does something different from the chatbots you know. It does not write. It chooses. Here is what that means, why speed is the selling point, and what to watch in Sarvam’s claims.
Quick Facts: Sarvam Prism
- What it is: Prism, an AI “decision model” from Bengaluru startup Sarvam. It picks from a fixed list of answers instead of writing text.
- Speed: A median 44 milliseconds per answer on one Nvidia B200 GPU.
- Score: 71.8 on Sarvam’s three-index average, versus 68.0 for Perplexity Decider and 61.1 for Microsoft-Decision-1.
- Access: Open weights under Apache 2.0, plus Sarvam’s API.
- Caveat: All numbers are Sarvam’s own. Independent tests have not been reported.
What Is Sarvam Prism?
Most large language models answer by generating text one token at a time. Prism skips that. You hand it a context, a question and a list of possible answers. It returns a probability for each option in a single pass, and it always lands on one of the choices you gave it.
Sarvam calls it a “System One decision model.” The label borrows from psychologist Daniel Kahneman, who split thinking into fast, intuitive judgment (System One) and slow, deliberate reasoning (System Two). Prism is built for the first kind.
According to the model card on Hugging Face, Prism handles three question types:
- A choice between several options.
- A yes or no decision.
- A score on an ordered scale.
It can read up to 131,072 tokens of context per question. It is a LoRA fine-tune of Sarvam’s own 105-billion-parameter model, Sarvam 105B.
Why 44 Milliseconds Matters
Speed is the headline. Sarvam says Prism responds in a median 44 milliseconds on one Nvidia B200 GPU. The first version took a median 664 milliseconds on two B200 GPUs.
| Version | Setup | Median response |
|---|---|---|
| First version | Two B200 GPUs, 16-bit | 664 ms |
| Current Prism | One B200 GPU, FP8 weights | 44 ms |
Sarvam credits two changes: converting the weights to FP8 and changing how the GPU runs the model. That is roughly 15 times faster. The 95th-percentile latency is 420 milliseconds, so a small share of answers take longer.
Why does that matter? Because decision tasks sit in the middle of live products. Think of a customer message that must be routed before a human sees it, or a guardrail that must flag a prompt injection attempt before an agent acts. Slow checks break the experience. Fast ones disappear.
What Can You Use It For?
- Routing: send a customer query to the right team or the right model.
- Security: flag prompt injection attempts.
- Quality checks: judge whether an AI agent completed its task.
- Voice and documents: Sarvam also tested speech-to-speech, trace evaluations and document intelligence.
Sarvam says it is sharing early benchmarks and will release more over time. For another look at AI in live products, see our guide to voice AI agents.
How Prism Compares, by Sarvam’s Numbers
Sarvam averaged scores across three indices: the Jev Decision Index (37 public benchmarks), its own recreation of the Artificial Analysis decision index, and an internal Indian-language benchmark. It tested 25 models.
| Model | Average score |
|---|---|
| Sarvam Prism | 71.8 |
| Perplexity Decider v1.2 | 68.0 |
| Jev 1.14 | 64.5 |
| Clef | 62.6 |
| Microsoft-Decision-1 | 61.1 |
Sarvam calls Prism the highest-scoring model under 100 milliseconds among those it tested. Two details deserve attention.
- The lead is mostly Indian languages. There, Prism scored 82.5 against 75.0 for Perplexity Decider and 70.8 for Jev 1.14.
- It did not win everywhere. On the Jev Decision Index, Perplexity Decider scored slightly higher: 67.7 against 67.0.
Calibration: Does the Model Know When It Is Wrong?
Sarvam also stresses calibration, meaning how closely a model’s stated confidence matches how often it is right. For a decision model, that is the trust test.
| Test (90%+ confidence) | Prism predicted errors | Prism actual errors |
|---|---|---|
| Indian languages | 5.4% | 4.4% |
| English | 3.3% | 2.2% |
Sarvam says other models made 2.7 to 4.8 times as many errors as their confidence predicted in Indian languages, with Jev 1.14 at the high end. In English, it says Clef came close (2.8% predicted, 2.3% actual), while Perplexity Decider, Microsoft-Decision-1 and Jev 1.14 underestimated their errors by factors of 1.5, 1.8 and 2.7.
The company says Prism made fewer errors than its confidence predicted at every threshold from 50% to 95%. It tested English, 10 Indian languages and Hinglish.
The Option-Order Fix
Decision models can have a quirk: shuffle the answer options and the answer may change. Across 2,000 questions, Sarvam says Prism changed 4.0% of its answers when options were shuffled. Fixing the internal order of options before scoring brought that to zero, with no extra computation or loss of accuracy, according to Sarvam.
Open Weights, With a Catch
Sarvam released Prism’s weights under the Apache 2.0 licence on Hugging Face, so enterprises can train and run it on their own infrastructure. It is also available through Sarvam’s API.
| Requirement | Detail |
|---|---|
| GPU | One with about 120 GB free memory, such as a B200 or H200 |
| Disk | About 106 GB |
| Software | Sarvam’s own inference code |
The catch: a custom FP8 weight layout means standard Hugging Face tools and general-purpose inference servers cannot load it.
Who Is Behind Sarvam?
Sarvam was founded by Vivek Raghavan and Pratyush Kumar, both formerly of IIT Madras’ AI4Bharat initiative. Earlier this year it introduced open-source models with 30 billion and 105 billion parameters.
- June 2026: raised $234 million in a Series B at a $1.5 billion valuation. HCLTech led with $150 million, with Bessemer Venture Partners joining Khosla Ventures and Peak XV Partners.
- Before that: $41 million across seed and Series A rounds.
- July 2026: HCLTech said it would set up its first AI data centre at the Odisha Sovereign AI Park in Bhubaneswar, with Sarvam and the Odisha government. The planned outlay is Rs 14,257 crore, including state assistance.
What to Watch Next
- Independent benchmarks that test Sarvam’s claims.
- More results from Sarvam’s promised benchmark releases.
- Whether enterprises adopt Prism despite the custom inference requirement.
- Responses from Perplexity, Microsoft and Jev’s maker, TypeSafe AI.
- How the decision-model category grows. For related AI model news, read our report on Google’s Gemini 4 Argon.
This article is for information only. All performance figures are Sarvam’s own and have not been independently verified.
Official Sources and Further Reading
See the Prism model card on Hugging Face. Browse more in our AI coverage and technology coverage.
About the Author and Our Editorial Standards
Laura Anderson covers technology, AI and business for Welp Magazine, part of Fair Media Group. We rely on primary reporting, label company claims, and update articles as facts change. This article was last updated on November 5, 2026.
How We Reported This
This article is based on reporting by Indian Startup News (November 4, 2026) and the Prism model card on Hugging Face. Benchmark figures, speed and calibration results come from Sarvam and are labeled as such. We have not run independent tests.
References and Further Reading
- Hugging Face: sarvamai/prism model card
- Indian Startup News: Sarvam launches an AI decision model
- Indian Startup News: Sarvam becomes India’s latest AI unicorn
Frequently Asked Questions About Sarvam Prism
What is Sarvam Prism?
Prism is an AI decision model from Sarvam. Given a context, a question and a list of possible answers, it returns a probability for each option in a single pass. It always picks from the options it is given.
How fast is Sarvam Prism?
Sarvam says Prism answers in a median 44 milliseconds on a single Nvidia B200 GPU. The 95th-percentile latency is 420 milliseconds. The first version took a median 664 milliseconds on two B200 GPUs.
What is a System One decision model?
The term borrows from Daniel Kahneman’s split between fast, intuitive judgment (System One) and slow, deliberate reasoning (System Two). A decision model is built for quick, structured choices, not long written answers. Sarvam credits TypeSafe AI’s Jev model, launched in September, with creating the category.
Is Sarvam Prism better than Microsoft and Perplexity?
By Sarvam’s own tests, it scores higher on average: 71.8 against 68.0 for Perplexity Decider v1.2 and 61.1 for Microsoft-Decision-1. But Perplexity Decider scored slightly higher on the Jev Decision Index, 67.7 against 67.0. The lead came largely from Indian languages.
What can Prism be used for?
Sarvam points to routing customer queries, flagging prompt injection attempts and judging whether an AI agent finished its task. It has also tested speech-to-speech, model routing and document intelligence.
Is Sarvam Prism open source?
The weights are released under the Apache 2.0 licence on Hugging Face, and the model is also available through Sarvam’s API. Standard Hugging Face tools cannot load it because of a custom FP8 weight layout, so teams need Sarvam’s own inference code.
What hardware does Prism need?
One GPU with about 120 GB of free memory, such as a B200 or H200, plus about 106 GB of disk space, according to Sarvam.
Who founded Sarvam?
Vivek Raghavan and Pratyush Kumar, who previously worked at IIT Madras’ AI4Bharat initiative. In June, Sarvam raised $234 million in a Series B at a $1.5 billion valuation, led by HCLTech with $150 million.
