Bengaluru-based artificial intelligence startup Sarvam AI has launched Prism, a decision-making model built on its Sarvam 105B architecture, designed to help enterprises automate tasks requiring fast responses and confidence-based predictions.
The company said Prism can deliver responses in 44 milliseconds while achieving an average score of 71.8 across three evaluation benchmarks. The model is intended for enterprise applications including customer support, document processing, cybersecurity and AI agent management.
Prism is a Low-Rank Adaptation (LoRA) fine-tuned version of Sarvam 105B. Unlike conventional large language models that generate responses sequentially, the new system processes inputs in a single forward pass and assigns probability scores to available options.
This approach allows organisations to use the model for classification and decision-making tasks without generating lengthy conversational responses.
According to Sarvam AI, Prism was evaluated against 24 other decision models on a single NVIDIA B200 GPU. The company reported that it achieved the highest average score among models operating with response times below 100 milliseconds.
Across three benchmarks, Prism recorded a score of 67 on the Jev Decision Index, 65.9 on the Artificial Analysis benchmark and 82.5 on an internal Indian-language evaluation.
The company has also developed the model to provide calibrated confidence scores, allowing businesses to establish thresholds for automated decisions and identify cases requiring human review.
Sarvam reported an error rate of 2.2% on English-language questions and 4.4% on Indian-language questions when the model's confidence level was at least 90%.
The company also assessed Prism's performance in AI agent monitoring, where the model achieved 86.4% accuracy in identifying agent activities requiring human review and 75% accuracy in determining whether tasks had been completed successfully.
In cybersecurity evaluations involving 5,517 questions across nine tasks, Prism recorded 89% accuracy. The assessments covered phishing detection, attempted jailbreaks and cybersecurity attack classification.
Sarvam said the model can evaluate an entire AI agent activity trace in approximately half a second on a single GPU and process inputs containing up to 131,072 tokens per question.
The company has released Prism's model weights on Hugging Face under the Apache 2.0 licence, allowing developers and enterprises to fine-tune the technology using their own datasets.
Prism is also accessible through Sarvam's application programming interface (API) and can run in FP8 precision on a single GPU with at least 120 GB of memory.
The launch expands Sarvam AI's portfolio beyond general-purpose language models into specialised systems designed for operational decision-making.
Its intended applications include routing customer queries, selecting AI models for particular tasks, analysing documents and reviewing automated agent activity.
While the reported benchmark results indicate the model's performance under testing conditions, its effectiveness across different enterprise environments will depend on specific workloads and deployment requirements.
Disclaimer: This article may include information derived from interviews, press releases, public statements, research, company communications and other publicly available or third-party sources. Such material may be summarised, paraphrased or contextualised for journalistic and editorial purposes. All rights in third-party content remain with their respective owners.