Startups & Technology

Decision models surge as startups focus on customer data

Decision models surge as startups focus on customer data

Decision models are tuned to draw on general intelligence to select an option and register confidence in that choice. Their speed and cost can make them useful in production, where a quick decision may matter more than the reasoning of a slower, more expensive frontier model.

TypeSafe’s Jev family continues to dominate usage in metrics visible through services such as OpenRouter, which processed more than 1.5 billion queries to TypeSafe models last week. The company became valued at $7.5 billion after releasing Jev. But the models are relatively easy to build, and dozens followed Jev’s launch. An Amazon engineer made one in a weekend; hyperscalers soon introduced their own offerings. On Tuesday, Microsoft CEO Satya Nadella said Microsoft-Decision-1.1 was rolling out for broad use on Azure.

The competition is also drawing startups. U.K.-based Sqwish announced a $3 million pre-seed round for its decision model, kettei fat. Cofounders Ushnish Sengupta and Federica Freddi met as Cambridge graduate students and had been building small LLMs before repurposing their architectures after Jev appeared.

Data may be a key differentiator. TypeSafe CEO Diogo Almedia told TechCrunch that calibrating Jev’s judgment with statistically adjusted synthetic data was the hardest part of development. Sqwish also uses synthetic data, but offers post-training tools that let customers use their own data to refine a model’s performance on tasks they care about. “They can keep uploading their production data,” Sengupta told TechCrunch, and their models can continue improving.

Benchmarks offer only a limited guide, builders say; developers can test models in their own settings. “The benchmark is more of an invitation to try the model out,” Sengupta said, with continual tuning giving customers a reason to stay. Microsoft engineering vice president Achint Srivastava likewise warned that “benchmaxed” models can fail in enterprise use. He said Microsoft-Decision-1.1 and Jev perform strongly across accuracy, reliability, speed and cost, including on unseen benchmarks designed to reflect real use cases.

Sqwish cofounder Freddi sees room for decision models—what she calls “System One” intelligence—to work alongside frontier LLMs, or “System Two” intelligence. She described the category as a potential $10 billion to $50 billion opportunity, even as frontier AI companies pursue valuations in the trillions.

Share

Comments (0)

Leave a comment

No comments yet. Be the first!