Sthanika Sieve 31B AI model secures second place on global Jev Decision Index

Hyderabad-based AI research lab Sthanika has taken second place globally on the Jev Decision Index with its specialized decision-making model, Sieve 31B, marking a significant milestone for India's indigenous artificial intelligence capabilities.

The open benchmark, hosted on Hugging Face, evaluates AI models built to make rapid, defined choices within complex software systems rather than generate conversational text. In the latest standings, Sieve 31B achieved a combined full score of 63.8 across public test suites and private evaluation sets maintained by benchmark administrators, placing second on the public suite with a score of 62.5.

The achievement places Sthanika's Indian-developed model ahead of prominent global contenders, including Jev and Perplexity Decider v1.1.

What sets Sieve 31B apart

While most public attention in artificial intelligence focuses on large language models built to write prose or summarize documents, Sthanika designed Sieve 31B to handle routine, high-volume automated decision-making. Software agents frequently process thousands of choices daily. Routing these tasks through massive general-purpose language models creates substantial computational overhead and operational costs.

Sieve 31B addresses this bottleneck by evaluating presented choices and delivering a defined answer alongside a verifiable confidence score. This targeted design allows enterprise applications to make fast, reliable judgements at a fraction of the cost required by traditional generative models.

The Sieve family also features smaller variants, including Sieve 2B Plus, Sieve 4B, and Sieve 9B Plus, which ranked near the top of their respective size categories. The architecture enables developers to deploy smaller models on local infrastructure for high-volume, privacy-sensitive workflows while scaling to larger models when task complexity demands it.

Indigenous innovation on the global stage

Sthanika operates as the applied AI research arm of the Hyderabad-headquartered PurpleTalk group. Its name is derived from the Sanskrit word for "local", reflecting a broader commitment to building sovereign, specialized AI capabilities in India.

The underlying decision framework is already integrated into PurpleTalk products such as Ello and Mina, where it handles routine evaluations in customer experience and workflow automation.

"Sieve 31B shows that a focused model fine-tuned in India can perform among the best in the world at this work," said Sridhar Muppidi, co-founder of PurpleTalk. He emphasised that the model's strong performance on private, unannounced test sets demonstrated real-world generalisation beyond public benchmark datasets.

PurpleTalk chief executive officer and co-founder Bharat Lingam noted that decision models will underpin a new generation of autonomous agentic systems across fields like voice processing, retail, and robotics. The lab is making its evaluation artifacts and model weights open to the global research community to support reproducible testing.

Key performance figures on the Jev Decision Index

The result underscores how Indian research institutions are advancing from technology adopters to global leaders in foundational model development, providing robust, economical AI tools built for real-world enterprise demands.