Sthānika, an applied artificial intelligence research lab under the PurpleTalk group (https://www.purpletalk.com/), announced today that its Sieve 31B model has achieved second place globally on the Jev Decision Index. The Jev Decision Index is an open benchmark hosted on Hugging Face, designed to evaluate AI models engineered for rapid decision-making.
The Sthānika model surpassed several international competitors, including the Jev model itself and Perplexity Decider v1.1, in the rigorous benchmarking process.
Sthānika: Performance Metrics and Model Specialization
Sieve 31B registered a Full score of 63.8, a figure derived from a combination of public test suites and private evaluations conducted by the benchmark’s maintainers. For public tests specifically, the model secured a score of 62.5, placing it second in that category.
Sthānika developed Sieve 31B as a cost-efficient decision model intended for integration into larger AI agent systems. Unlike general-purpose large language models (LLMs) that are often tasked with both conversational and routine decision-making processes, Sieve 31B is optimized for making focused decisions. This specialization allows it to handle repetitive decision tasks more reliably and at a lower computational cost compared to LLMs, which are primarily tuned for human interaction and language generation. The cost efficiencies become increasingly relevant as AI agents assume more operational responsibilities.
The broader Sieve model family also demonstrated strong performance, with Sieve 2B Plus, Sieve 4B, and Sieve 9B Plus each ranking among the top models within their respective size categories on the same leaderboard. Developers can deploy these smaller models on on-premise GPUs for tasks requiring data privacy and high-volume processing, while reserving larger models for more complex computational demands.
Industry Impact and Future Outlook
While much of the public discourse surrounding AI centers on models capable of writing, summarizing, and conversing, a significant portion of AI applications within software systems involves less visible, repetitive decision-making. Sieve models are engineered for this specific function, interpreting questions and available choices to provide defined answers with an accompanying confidence score. This capability is critical for applications where consistency, speed, and auditability are as important as accuracy.
Sridhar Muppidi, co-founder of PurpleTalk, commented on the model’s performance, stating, “Sieve 31B demonstrates that a focused model developed in India can compete globally in this domain. The private tests, which involve diverse domains beyond the public suite, were particularly significant, as Sieve showed strong generalization capabilities.”
The Sieve technology builds upon Sthānika’s prior work in developing compact, confidence-aware decision models, which are already integrated into PurpleTalk’s existing products, Ello and Mina.
Bharat Lingam, CEO and co-founder of PurpleTalk, indicated that decision models are poised to enable a new category of agentic systems. “We anticipate these systems will offer enhanced reliability, granular control, and low latency crucial for applications such as voice interfaces, customer experience platforms, and robotics,” Lingam said, inviting the research community to test and build upon their published results.
For those interested in the latest AI news and other industry developments, you can find more information on TechnoSports (https://technosports.co.in/). Evaluation artifacts and models are publicly available via Hugging Face collections, and full results can be accessed on the Jev Decision Index.
