Sarvam AI, the Bengaluru-based AI startup, has introduced Sarvam Prism, a decision making model designed to help businesses make faster and more reliable decisions using AI. The model can process decisions in just 44 milliseconds, making it suitable for enterprise applications where AI systems need to respond quickly without generating lengthy answers.
Built on Sarvam 105B, Prism is designed to handle repeated decision-making tasks, such as identifying customer queries that require escalation, assessing security alerts and determining when an AI agent’s work needs human intervention.
Read More | Sarvam Signs Five-Year MoU with Defence Ministry to Build AI for Army, Navy and Air Force
According to Sarvam AI, the model achieves a score of 71.8 on the average of three decision-focused benchmarks. The company reported this performance while running Prism on an NVIDIA B200 GPU. The model falls under what Sarvam calls a System One decision model, which focuses on making quick judgements rather than producing detailed, step-by-step responses.
Why has Sarvam AI developed Prism for enterprise decision-making?
AI systems are increasingly being used across business operations. However, not every task requires an AI model to generate a detailed response. Many enterprise applications need AI to make smaller decisions repeatedly and quickly.
For example, an AI-powered customer support system may need to decide whether a query can be handled automatically or should be passed to a human agent. Similarly, security systems may need to assess alerts and determine which ones require immediate attention.
These decisions influence how businesses manage customer interactions, allocate tasks between AI models and identify situations that need additional checks. Sarvam AI says it tested Prism against 24 other decision models using the same benchmarks. According to the company, Prism recorded the highest average score among models that delivered responses in under 100 milliseconds.
Read More | India Wants to Train AI in Space—Here’s What Pixxel & Sarvam Are Building
While some competing models achieved higher overall scores, they took between two and six times longer to respond. The company also highlighted Prism’s performance across Indian languages. On its internal Indian-language benchmark, the model scored 82.5, outperforming Perplexity Decider v1.2, Jev 1.14, Clef and Microsoft-Decision-1 in the comparison shared by Sarvam.
How reliable are Sarvam Prism’s AI decisions?
Speed is important for enterprise AI applications, but it is not the only factor. Businesses also need to understand how confident an AI model is in its decisions and whether that confidence reflects its actual accuracy. This is where calibration becomes important.
In simple terms, calibration measures whether an AI model’s confidence level matches how often its decisions are correct. For instance, if a model claims to be 90% confident in its decisions, its actual accuracy should broadly reflect that confidence.
According to Sarvam AI’s tests, Prism recorded fewer errors than its confidence scores predicted across confidence thresholds ranging from 50% to 95%. At a confidence level of 90% or higher, the model predicted an error rate of 3.3% in English but recorded an actual error rate of 2.2%.
For Indian languages, it predicted an error rate of 5.4%, while the recorded figure was 4.4%. Sarvam also tested whether Prism’s decisions remained consistent when the order of available answer options changed. After standardising the model’s internal ordering, it returned the same answers across all 2,000 tests, even when the options were shuffled.
This consistency is particularly relevant for businesses using AI to make repeated decisions, where changing the order of available choices should not influence the outcome.
How does Sarvam Prism achieve 44-millisecond response times?
Unlike conventional generative AI models that produce responses one token at a time, Sarvam Prism is designed to evaluate available choices and return their respective probabilities. It is a fine-tuned version of Sarvam 105B, developed specifically for decision-making tasks.
Sarvam AI says several engineering improvements helped reduce the model’s median response time from 664 milliseconds to 44 milliseconds.
These improvements include:
- FP8 weights: A numerical format that helps reduce the memory required to run the model.
- Single-GPU execution: The model is designed to run on a single GPU, simplifying its deployment.
- CUDA Graphs: A technology that helps reduce the overhead involved in executing GPU operations.
- Merged fine-tuning weights: An optimisation that combines model weights to improve execution efficiency.
How can developers access Sarvam Prism?
Sarvam AI has made Prism’s model weights available on Hugging Face under the Apache 2.0 licence. Developers can also access the model through the Sarvam API. This gives developers and businesses an opportunity to explore the model for applications where response speed, confidence scoring and consistency are important.
For enterprises exploring AI automation, Sarvam Prism offers an alternative to using general-purpose generative AI models for every task. Instead of generating lengthy responses, it focuses on making quick decisions that can help businesses route requests, assess risks and determine when human intervention is necessary.
Indrani Priyadarshini is a journalist and editorial professional specialising in technology, artificial intelligence, smart cities, green energy, and digital transformation. With over four years of experience in tech journalism and digital media, she is known for turning complex industry developments into clear, engaging, and insightful stories. Her expertise spans reporting, editorial strategy, digital publishing workflows, and in-depth coverage of emerging technologies shaping the future.
