44 milliseconds per answer: Sarvam launches an AI decision model faster than Microsoft and Perplexity

Sarvam introduces Prism, a rapid AI decision model featuring a 44ms response time, open weights, and capabilities aimed at enterprises.

Related: Enterprise AI startup Ema raises $77 million in Series B funding led by Creaegis

Sarvam AI, an artificial intelligence startup headquartered in Bengaluru, has unveiled Sarvam Prism, a new AI decision model built to make fast and dependable choices for enterprise workflows.

According to Sarvam’s benchmarks, Prism, which is built on the company’s Sarvam 105B foundation model, delivers a median response time of just 44 milliseconds on a single NVIDIA B200 GPU. The company evaluated 25 decision models and reported that Prism earned an average score of 71.8 across three decision-focused indices. Among the models in its comparison that responded in under 100 milliseconds, it was the highest-scoring one.

Unlike traditional AI chatbots that produce lengthy text responses, Prism concentrates on choosing the most suitable answer from a fixed set of options. This makes it valuable for tasks where software needs a quick decision instead of a detailed explanation.

The launch underscores Sarvam’s expanding focus on enterprise AI, where speed, consistency and confidence in automated decisions are essential.

Sarvam Prism is a purpose-built AI model for decision-making. It falls into a category Sarvam refers to as a System One decision model, drawing on the notion of quick, intuitive human judgment.

Conventional large language models produce responses one token at a time. Decision models, on the other hand, operate in a different way. They assess a set of possible choices and assign a probability to each option.

Take, for instance, a customer support platform that receives a message requesting a refund. Rather than generating a lengthy response, Prism can help sort the request into options like refund, replacement, or escalation.

This design makes the model well suited to repeated decisions within business software.

Main Capabilities of Sarvam Prism

• Quick decision-making: In Sarvam’s test environment, it delivers structured decisions with a median response time of 44 milliseconds.
• Confidence scoring: It offers probability estimates for the options that are available.
• Consistent outputs: It is built to minimize variations that result from reordering answer options.
• Enterprise customization: Companies can fine-tune its open weights with their own data.
• Indian-language support: It has been tested on an internal benchmark that covers Indian languages.

These features are intended to make Prism valuable in workflows where businesses require fast, repeatable, and measurable decisions.

A major highlight of the launch is how Prism performs against other decision models, such as Microsoft-Decision-1 and Perplexity Decider.

In Sarvam’s comparison of 25 models, Prism earned an average score of 71.8 across three indices. It also posted a median response time of 44 milliseconds on an NVIDIA B200 GPU.

Below is how several of the models compared in the results the company published:

The results indicate that Prism provides a solid blend of speed and accuracy within the group that was tested. Still, the comparison relies on Sarvam’s published methodology and should not be regarded as a universal ranking for every AI task.

In Which Areas Did Sarvam Prism Do Better?

On Sarvam’s internal benchmark for Indian languages, Prism achieved a score of 82.5. By contrast, Perplexity Decider v1.2 reached 75.0, and Jev 1.14 came in at 70.8.

Yet on the Jev Decision Index, Perplexity earned 67.7, a little higher than Prism’s 67.0. This indicates that a model’s performance may differ from one evaluation category to another.

All in all, Prism’s biggest edge in the published comparison came from combining a high average score with a low response time.

When businesses rely on AI for thousands of daily decisions, speed becomes a major factor. Even minor delays can impact customer experience and add processing time across large workflows.

For instance, an e-commerce platform might need to classify customer queries, spot suspicious transactions, and route support requests. These systems can handle requests with less waiting time when the decision model is faster.

According to Sarvam, Prism’s median response time is 44 milliseconds. That makes it especially relevant for applications needing frequent decisions with low latency.

In what ways did Sarvam make Prism faster?

Sarvam has outlined multiple technical enhancements that contributed to lowering Prism’s response time.

• FP8 precision: To lower memory requirements, the company converted the model weights to FP8.
• GPU optimisation: A grouped kernel was used to run all 128 experts efficiently.
• CUDA Graphs: The model’s forward pass was captured to cut down execution overhead.
• Weight optimisation: The fine-tuning adapter was merged into the model weights.

Per Sarvam, these changes cut median response time from 664 milliseconds to 44 milliseconds, with no loss of accuracy in its testing.

For enterprise AI, speed by itself is not sufficient. Companies also need to know whether a model’s confidence aligns with its real accuracy.

This concept is called calibration.

For instance, when an AI model claims 90% confidence in a decision, its actual error rate should roughly match what that confidence level implies.

Sarvam states that Prism scored well on this metric in both English and Indian-language tests.

At confidence levels of 90% or above, the model forecast a 3.3% error rate for English responses while its actual error rate was 2.2%. For Indian-language responses, it forecast 5.4% errors and recorded 4.4%.

Why Is Calibration Important for Businesses?

Dependable confidence estimates help businesses determine when to automate a task and when a human should be involved.

For example, a customer support system could automatically route a request whenever the model is confident. Yet it could forward uncertain or sensitive cases to a human employee.

This kind of approach can make AI systems more practical for business applications. It also assists organisations in building workflows with clear oversight.

Sarvam also examined whether altering the order of answer options influenced Prism’s decisions. The company reported that once the option order was standardised, the model delivered consistent results across 2,000 shuffled questions.

Sarvam built Prism to handle structured decision-making tasks in enterprise applications. Its quick response time and probability-based outputs may assist a range of business workflows.

1. Automation of Customer Support

Companies get a high number of customer inquiries each day. Prism assists in categorizing requests, determining their intent, and directing them to the appropriate department.

For instance, it could tell the difference between billing inquiries, refund requests, and technical complaints.

2. Evaluation of AI Agents

More and more, companies are relying on AI agents to carry out tasks. Still, they require methods to assess whether these agents have correctly followed instructions.

Prism is able to assist in evaluating agent outputs and to help pinpoint actions that call for human review.

3. Document Intelligence

Companies handle invoices, forms, contracts, and various other documents. A decision model can assist in categorizing documents, spotting pertinent details, and directing them to the right workflows.

4. Security and Risk Assessments

Prism is also able to assist with structured security decisions, including categorizing alerts and spotting activity that may be risky.

These represent possible uses of its capabilities. Real-world performance will hinge on how well it is implemented, the quality of training data, and the particular workflow.

Sarvam notes that initial applications cover customer support routing, prompt-injection guardrails, document intelligence, model routing, and AI-agent evaluation.

Another key aspect of the launch is that Prism’s model weights are available under the Apache 2.0 licence.

This provides enterprises with more flexibility to customise and deploy the model. Rather than relying solely on an external service, businesses can fine-tune the model with their own decision data and host it on their own infrastructure.

The model can also be accessed through Sarvam’s API.

Sarvam states that running Prism in FP8 precision needs a GPU with at least 120 GB of memory. This offers technical teams the option to deploy the model within their own infrastructure, depending on their hardware and operational requirements.

What Makes Open-Weight AI So Significant?

Open-weight models allow businesses to have greater control over the way they use AI.

The main advantages include:

• Customisation: Companies are able to tailor the model to suit their own business decisions.
• Data control: Organisations have the option to run the model inside their own infrastructure.
• Integration: Developers can link it up with the software and workflows they already use.
• Greater flexibility: Businesses can pick between self-hosting and supported API deployment.

For enterprises that deal with sensitive information, these choices may be particularly valuable.

The release of Sarvam Prism marks yet another milestone in India’s expanding artificial intelligence landscape.

Indian AI startups are developing models and applications tailored to local languages, enterprise automation, and specialised business needs. Sarvam’s newest release tackles a particular challenge: enabling AI decisions to be made faster and more reliably.

Its benchmark results for Indian languages are especially pertinent for companies serving diverse customer bases throughout the country.

In addition, the availability of open weights may prompt developers and enterprises to experiment with customised decision systems.

Yet the model’s long-term influence will hinge on real-world adoption, independent evaluation, and how well it performs across various industries.

The launch also points to a broader shift in how AI is being developed. Firms are looking past general-purpose chatbots and are instead exploring specialised models built for specific tasks.

Through Prism, Sarvam is aiming at businesses that require rapid, structured decisions backed by confidence levels.

The company’s next hurdle will be broadening adoption throughout enterprise workflows while proving the model’s performance beyond its first benchmarks.

Its emphasis on Indian languages, open weights and low-latency inference may help it draw businesses seeking alternatives to larger general-purpose AI systems.

Meanwhile, competition from Microsoft, Perplexity and other AI developers will keep driving improvements in speed, accuracy and reliability.

All in all, Sarvam Prism’s 44-millisecond response time makes it a noteworthy development in enterprise AI. Instead of replacing general-purpose chatbots, it serves a different need: helping software make fast, structured decisions with measurable confidence.