Sarvam Prism: New AI Model Hits Blazing 44ms Response Time
Sarvam AI introduces Prism, a high-speed decision model for enterprises with open weights and 44ms response times for business workflows.
In a significant leap for enterprise automation, the Indian artificial intelligence firm Sarvam AI has unveiled Sarvam Prism, a specialized decision-making model designed for extreme speed and precision. Capable of delivering responses in just 44 milliseconds, the model aims to transform how businesses handle structured decisions and complex workflows.
Built for Speed and Precision
Sarvam Prism is built upon the company's robust Sarvam 105B base model. It is specifically engineered to tackle tasks that require rapid, structured outputs, such as routing customer support queries, analyzing dense documents, and evaluating the performance of other AI agents.
The performance metrics are striking. When tested on an NVIDIA B200 GPU, Prism clocked an average response time of 44 milliseconds. More impressively, it achieved an average score of 71.8 across three key benchmarks. According to Sarvam, this is the highest score recorded among 25 different decision models tested that could respond in under 100 milliseconds.
Dominating Indian Language Benchmarks
Sarvam Prism isn't just fast; it is highly effective in the Indian linguistic landscape. In internal benchmarks focusing on Indian languages, Prism scored 82.5, significantly outperforming competitors like Perplexity Decider v1.2 (which scored 75) and Jev 1.14 (which scored 70.8).
Reliability and Confidence
For enterprises, the biggest hurdle in AI automation is trust. Sarvam has addressed this by designing Prism to provide reliable confidence estimates. The company's testing revealed that the model is often more accurate than it predicts:
- For responses delivered with at least 90% confidence, the error rate was only 2.2% for English.
- For Indian languages, the error rate sat at 4.4%.
Sarvam noted that Prism consistently made fewer errors than its confidence scores predicted across all tested thresholds, ranging from 50% to 95%.
Consistency and Security
One common flaw in AI models is "positional bias," where the answer changes based on the order of the options provided. Sarvam claims to have solved this; Prism's answers remained identical across 2,000 questions, regardless of how the options were arranged.
The model also shows immense promise in the realm of cybersecurity. On an internal benchmark consisting of 5,517 questions, Prism achieved 89% accuracy, beating out Clef (84.5%) and Jev 1.14 (78.6%).
Open Weights for Enterprise Control
In a move to empower businesses, Sarvam has released the weights of Prism under the Apache 2.0 licence. This "open weights" approach allows enterprises to:
- Train the model on their own proprietary data.
- Deploy and run the model on their own private infrastructure for better security and control.
- Access the model via Sarvam's API for quicker integration.
Primary Use Cases
The company highlighted several early applications where Prism can be deployed immediately to improve efficiency:
- Customer Support Routing: Instantly directing users to the right department.
- Document Intelligence: Rapidly extracting and analyzing structured data from files.
- Security Checks: Enhancing automated threat detection and verification.
- AI-Agent Evaluation: Acting as a judge to monitor and score the performance of other AI agents.