Little-Known Indian Lab Sthānika Claims Its 31B Model Outsmarts Perplexity

Little-known Indian lab Sthānika said its Sieve 31B decision model ranked second on the Jev Decision Index, ahead of Perplexity, in a bid to win over enterprise buyers.

Key Takeaways

• Sthānika said its 31 billion-parameter model Sieve 31B ranked second on the Jev Decision Index
• Sthānika said Sieve 31B beat most international decision models, including Jev and Perplexity
• The Jev Decision Index evaluates decision-oriented model performance rather than casual chat quality
• Next disclosures of reproducibility and access, such as API availability or open weights, will matter to developers and buyers

Sthānika said Sieve 31B beat most international decision models, including Jev and Perplexity. The company is making a positioning play as much as a technical one, because leaderboard placement is one of the fastest ways a smaller lab can win attention from developers, buyers and model evaluators.

Benchmark Rank Is Marketing And Signal

The release centers on the Jev Decision Index, a benchmark aimed at evaluating decision-oriented model performance rather than casual chat quality.

Many enterprise buyers prioritize whether a model can follow constraints, weigh options and complete stepwise tasks over conversational flair.

Sieve 31B refers to the model scale. Thirty-one billion parameters is large enough to compete in serious open-model discussions, while still smaller than some frontier proprietary systems.

Parameters are adjustable weights learned during training. They are not intelligence by themselves, but they usually correlate with the amount of pattern capacity a model can hold.

Sthānika reported the ranking as proof that an Indian-built model can compete with better-funded international peers.

That pitch lands at a moment when regional labs are trying to show they can do more than fine-tune U.S. or Chinese systems.

How Ranking Affects Sales Conversations

A high score on a decision benchmark can open sales doors. It can earn a model pilot request long before a full independent review appears, especially in categories like claims handling, procurement or operations, where models must compare options and enforce policy constraints.

A smaller lab can use one benchmark win to enter the enterprise sales funnel without needing the largest model family.

It only needs to show that, on one decision-heavy workload, it belongs on shortlists.

Also Read: Zai’s GLM-5.3 Hits 160, Shrinks Open Models’ Cyber Gap To 6 Months

From National AI Ambition To Product Claims

The announcement also fits a wider shift in the global AI market. Early attention clustered around a few U.S. labs with giant budgets and cloud backing.

The next phase is more fragmented. Regional labs are trying to win in language coverage, local enterprise sales and narrower performance domains.

India is an important proving ground because it offers a large developer base, strong English usage and growing interest in sovereign AI capacity.

A strong model result from an Indian lab therefore carries two messages at once. One is technical credibility and the other is geographic self-confidence.

The hard part comes after the headline.

Benchmark success does not guarantee production success. Buyers still need latency data, pricing, hosting options and evidence that the model behaves consistently on real workflows.

What Sthānika Must Show Next

The next disclosures that matter are reproducibility and access.

Developers will want to know whether Sieve 31B is available through an API, with open weights or only in limited trials. Enterprise buyers will want task-level comparisons, not just a leaderboard position.

If Sthānika can turn this score into pilots, the ranking becomes commercially useful.

If not, it will remain what many benchmark wins become, a sharp marketing moment with limited staying power.

How A Decision Benchmark Differs From Chat Tests

Sieve 31B is being framed as a decision model, which usually means the benchmark asks it to choose among options under rules, not just produce fluent text. A model like Perplexity may excel at retrieval and answer presentation, while a decision benchmark cares more about structured reasoning.

For Sthānika, the useful next evidence is a breakdown showing which task families pushed Sieve 31B to No. 2 and whether outside developers can reproduce the result.

Without that, buyers only know the headline number, not the conditions behind it.

Read Next: Agent Memory Overtakes Full Context With 16 Tokens In Real FactKG Test