Sthānika's Sieve 31B Beats Jev, Perplexity Models in Global AI Benchmark

Hyderabad-based artificial intelligence research lab Sthānika, part of the PurpleTalk group, has announced that its Sieve 31B model secured second position globally on the Jev Decision Index, an open benchmark hosted on Hugging Face that evaluates AI models designed for rapid decision-making.

According to the company, Sieve 31B recorded a full score of 63.8, combining results from public evaluations and private tests conducted by the benchmark's maintainers. The model also secured second position on the public test suite with a score of 62.5.

Sthānika stated that its model ranked ahead of several international decision-making models, including Jev and Perplexity Decider v1.1.

Unlike general-purpose large language models (LLMs), which are primarily designed to generate text, answer questions and conduct conversations, Sieve 31B focuses on selecting appropriate responses from defined options.

The model processes a question and its available choices before returning a decision accompanied by a confidence score. This allows software systems to assess and act on its outputs.

According to Sthānika, specialised decision models can reduce the computational costs associated with repeatedly using general-purpose LLMs for routine tasks within AI agents.

The company has also developed smaller models within the Sieve family, including Sieve 2B Plus, Sieve 4B and Sieve 9B Plus. These models have secured positions among the leading performers in their respective size categories on the same benchmark.

The smaller variants are designed to operate on on-premise graphics processing units (GPUs), allowing organisations to manage routine workloads locally while using larger models for more complex decisions.

Commenting on the development, Sridhar Muppidi, Co-founder, PurpleTalk, said, "Sieve 31B shows that a focused model fine-tuned in India can perform among the best in the world at this work."

He added that the model's performance on private tests was particularly significant because these evaluations covered domains outside the publicly available test suite.

Sthānika's earlier decision-making models are already deployed within PurpleTalk products, including Ello and Mina, where they support routine decision-making processes.

The company is exploring further applications for specialised decision models in voice-based systems, customer experience platforms and robotics, where response speed, consistency and reliability are important operational requirements.

Bharat Lingam, CEO and Co-founder, PurpleTalk, said, "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."

He also stated that the company intends to publish the materials required to reproduce its evaluation results and encourage independent testing by the research community.

Sthānika's latest benchmark performance reflects its focus on developing specialised AI models for decision-making tasks, alongside the wider industry adoption of agentic AI systems.