Sieve 31B recorded a Full score of 63.8, combining the public test suite with private tests conducted by the benchmark’s maintainers. On the public tests, the model ranked second with a full score of 62.5.
Sthanika developed Sieve 31B as a cost-efficient decision model that AI agents can call as one component of a larger system. According to the lab, it makes decisions more reliably than general-purpose large language models (LLMs).
Many AI agents currently route every task to an LLM, including planning and routine decisions that a dedicated decision model can handle at a lower cost. This expense increases as agents take on more tasks previously performed by people. General-purpose LLMs are designed to converse with people, making them suitable for language-related tasks but expensive for selecting between defined options thousands of times a day. In comparison, decision models are extremely inexpensive for such tasks.
The wider Sieve family includes Sieve 2B Plus, Sieve 4B and Sieve 9B Plus, each of which also placed among the top models in its respective size class on the same leaderboard. Developers can run smaller models on their own on-premise GPUs for everyday tasks requiring privacy and high-volume processing, while switching to larger, more powerful models when applications demand more advanced capabilities.
While much of the public discussion around AI focuses on models that write, summarise and converse, a substantial part of AI operating inside software handles less visible tasks: making small decisions thousands of times a day. Sieve is designed specifically for this decision-making work.
The model processes a question and the available choices as presented, then returns a defined answer accompanied by a confidence score that software can inspect and act upon. This makes it suitable for applications in which consistency, speed and the ability to audit individual decisions are as important as accuracy.
“Sieve 31B shows that a focused model fine-tuned in India can perform among the best in the world at this work. The private tests matter most to us. They come from domains outside the public suite, and Sieve generalised better than most of the field. It learned the task, not the test,” said Sridhar Muppidi, Co-founder of PurpleTalk.
Sieve builds on Sthanika’s earlier work on compact, confidence-aware decision models, which are already deployed within PurpleTalk products, including Ello and Mina. In these products, the models handle routine judgements faster and more economically than general-purpose generative models.
The lab also expects decision models to enable a new category of agentic systems offering greater reliability, more granular control and low latency in areas such as voice, customer experience and robotics.
“We believe decision models will open up a new category of agentic systems: more reliable, more controllable and fast enough for voice, customer experience and robotics. We are publishing everything needed to reproduce our results, and we invite the research community to test it, challenge it and build on it,” said Bharat Lingam, CEO & Co-founder, PurpleTalk.
The models and evaluation artifacts are available through Sthanika’s collection on Hugging Face at https://huggingface.co/collections/sthanika-ai/sieve. The full evaluation results can be accessed through the Jev Decision Index at https://huggingface.co/spaces/multimodalart/jev-decision-index.
Sthanika is a Hyderabad-based AI research lab that develops open-weight models for problems that general-purpose models were not designed to solve. Its name is derived from the Sanskrit word for “local.” The lab is part of the PurpleTalk group.
Founded in 2008, PurpleTalk builds AI companies and services. The group has worked across sectors including healthcare, agriculture, financial services and retail. Its portfolio includes (x)cube LABS, Sthanika, Ello and Mina.
About The Author
Comment List