More than a quarter of requests sent through Vercel’s AI model router in October went to decision models. The type of model had been introduced just one month earlier, but quickly gained ground in tasks where a fast, inexpensive choice can matter more than a larger model’s detailed reasoning.
Decision models are built on large language models but do not generate output tokens. Instead, they choose among preset options and report how confident they are in the decision. They can be useful in production applications where latency and cost are particularly important.
TypeSafe became a $7.5 billion company after releasing its Jev family of models. Available usage figures suggest Jev remains a leading option: the OpenRouter service handled more than 1.5 billion queries to TypeSafe models in the previous week.
More companies enter the market
Decision models are relatively easy to build, and dozens of alternatives appeared after Jev’s release. An Amazon engineer made one over a weekend, and major cloud providers soon developed their own options. On Tuesday, Microsoft Chief Executive Officer Satya Nadella announced that the company was making Microsoft-Decision-1.1 immediately available for broad use on its Azure platform.
Among smaller companies, U.K.-based Sqwish announced a $3 million pre-seed funding round. The company, which develops a decision model called kettei fat, was founded by Ushnish Sengupta and Federica Freddi, who met as postgraduate students at Cambridge. They were working on small language models before Jev appeared and realized their architectures could also be repurposed for a decision model.
The role of data and practical tests
Data plays an important part in making the models successful. TypeSafe Chief Executive Officer Diogo Almeida told TechCrunch that the most difficult part of building Jev was finely calibrating its judgment with statistically adjusted synthetic data. Sqwish also uses synthetic data, but aims to differentiate itself with built-in post-training tools that let customers improve the model’s decisions using their own data.
Sqwish Chief Executive Officer Ushnish Sengupta said customers can continually upload data from production, allowing the model to improve at the tasks that matter to them. Benchmark results, he said, can encourage developers to try a model, while continual tuning gives them a reason to keep using it.
Ranking decision models is difficult, however. Although kettei fat has performed well on some benchmarks, several developers of the technology say the real test is performance in a particular application. Developers can generally swap one model for another and compare the results.
Microsoft also sought to balance cost, latency, capability and calibration when developing Microsoft-Decision-1.1. Achint Srivastava, the engineering vice president who led the project, said placing too much emphasis on benchmarks can be misleading: models optimized for those tests can quickly perform poorly in real enterprise use. According to Microsoft’s testing, Microsoft-Decision-1.1 and Jev are among the few models that perform strongly across accuracy, reliability, speed and cost efficiency, including on previously unseen tests reflecting real-world use cases.
A distinct market for fast decisions
Federica Freddi expects future systems to combine fast, so-called System One intelligence, such as decision models, with System Two intelligence represented by the most advanced large language models.
Freddi believes the market for fast decisions could itself represent a $10 billion to $50 billion opportunity. In her view, it is a central business goal for a startup such as Sqwish, while it may remain a side project for the largest AI companies.



