Every AI agent you will ever meet spends its day making tiny decisions. Which tool should it call next? Is this claim true, or a hallucination? Should this ticket go to billing, or to support? The giant chat models that power most agents can answer those questions — slowly, and at a price. In Hyderabad, a small research lab asked a different question: what if a model did nothing but decide, and did it better than anyone else on Earth?
The answer is Sieve, the family of decision models from Sthānika AI, the research lab of Hyderabad’s PurpleTalk group. Give Sieve a piece of text or a JSON record and a question with a fixed set of options, and it returns a probability for every option — no free text, no sampling, nothing to parse. And on the Jev Decision Index, the open benchmark hosted on Hugging Face that measures exactly this skill, the lab’s flagship Sieve-31B now scores 71.1, tied for first place and ahead of Jev 1.14 itself at 60.24. When the lab announced its first result on the same board in October, Sieve-31B placed second in the world with a full score of 63.8, ahead of both Jev and Perplexity’s Decider.
The trick is specialisation. A decision model does not write essays or hold conversations. It reads a situation, weighs a fixed set of options, and answers with numbers a piece of software can act on — which department, how urgent, approve or escalate, true or false. Sieve scores each option as its own branch, so shuffling the order of the options can never change the answer. And the lab calibrates its confidence scores with unusual care: the company says Sieve-4B carries a calibration error of just 3.1 points, and only 0.5% of the answers it gives with at least 95% confidence turn out to be wrong. Speed comes with it — Sieve-9B-v2 answers in 50.5 milliseconds, the lab says, more than twice as fast as its bigger sibling, in about 20 GB of GPU memory.
Sieve runs from 0.8 billion to 31 billion parameters, so there is a size for every job. Sieve-2B-v3 handles lightweight intent detection and routing, scoring 81.7 macro-F1 on the CLINC150 benchmark. Sieve-27B-v2 is the family’s most accurate model on intent detection, claim verification and checks for AI hallucination, running on a single 80 GB GPU. And the flagship Sieve-31B, a fine-tuned adapter on Google’s Gemma 4 31B, is the one leading the global board — scoring 75.3 on private tasks from domains it has never seen, the lab says, against 68.3 on the skills it was trained for. The lab even publishes its compute: Sieve-31B trained in 12.03 hours on two B200 GPUs, and the 9B model in 5.56 hours on two H100s. A small team, a short training run, and a world-leading score.
The name gives away the mission. Sthānika is Sanskrit for “local” — of this place, as the lab puts it — and the lab builds for the place it serves. Its Indic crop-advisory model answers Indian farmers’ questions in English and 10 Indian languages, in the style of Kisan Call Centre responses, and the lab says it beats its own 12-billion-parameter base by 1.14 points and edges past a 27-billion-parameter model with more than twice its size. The same instinct runs through the lab’s public research: a 500-question probe of what AI models actually know about India, a benchmark of 16 speech-to-text models across 22 Indian languages measured on real audio, a study of how much ability AI loses on romanised Hindi, and a measurement of how many tokens Telugu costs on every major tokenizer. Every claim traces to a published benchmark, and every model ships with its weights, its dataset and its evaluation set released openly.
The work already ships in the real world. The lab says its earlier decision models run inside PurpleTalk products Ello and Mina, making routine judgements faster and more economically than general-purpose models. PurpleTalk, founded in 2008 and led by CEO and co-founder Bharat Lingam with co-founder Sridhar Muppidi, builds AI companies and services across healthcare, agriculture, financial services and retail — and Sthānika is its research arm, publishing the small, specialist models its products run on. The Sieve weights and code sit openly on Hugging Face and GitHub for anyone to download, run on their own machines, and build on.
More Read
It is a quietly radical idea of what Indian AI can be: not a bigger model chasing the frontier, but a smaller one aimed precisely at the job, trained in hours, measured in public, and good enough to sit at the top of a world leaderboard. From a Hyderabad lab, built for India first — and now deciding, thousands of times a second, for anyone who wants to use it.
Keep reading: The Kitchen That Cooks Itself: Inside upliance.ai, Bengaluru’s AI Chef for Every Indian Home · The Chef That Never Gets Tired: Inside Nosh, the Bengaluru Startup Building India’s AI Robo-Chef · More Industrial Automation & Technology stories
Like stories like this? Add Exceptional Technology as a preferred source on Google to see more of our reporting in your search results and Discover feed.
