For four years the entire industry has been optimising for one thing: language. Bigger models, longer context, more fluent prose. This week two small releases pointed in the opposite direction, and developers noticed immediately. Both models share a strange feature for the age of the chatbot. They do not produce words at all.
The louder of the two is Jev, from a startup called TypeSafe AI. Its founder, Diogo Almeida, helped build ChatGPT and co-invented reinforcement learning from human feedback, the training method behind much of the current boom. He left OpenAI disillusioned. "We have lightning in a bottle, and yet it is not useful," he told TechCrunch. His diagnosis: we optimised for human language, but computers speak a different language, and automation needs decisions, not paragraphs.
So Jev does not write. Given a task and a fixed set of options, it returns a probability for each: what TypeSafe calls a calibrated decision. Because the possible outputs are defined in advance, the model cannot hallucinate a new one. Its output tokens are free, and input is metered by the billion rather than the million. Demand was heavy enough that the company briefly lost the ability to serve its own API. One engineer at Vercel swapped a safety classifier from an OpenAI model to Jev and reported results five to eighteen times faster. Another tester found a rival model slightly more accurate but ten to twenty times more expensive, and, crucially, unwilling to hand back a real confidence score.
That last point is the interesting one. A probability is honest about its own uncertainty. As one developer put it, if the answer comes back at 50 percent you can treat it as a coin toss and ignore it; at 95 percent you can act. The model delegates the hard judgement back to the user rather than dressing a guess in confident prose.
Four days after Jev, a small firm called Cua open-sourced its own version of the idea. CUA-S1-FORMS is tiny, 706,048 parameters and a 2.8 megabyte file, trained to do one bounded job: decide how to fill a field on a form. It too scores options instead of writing text, and Cua released the weights, the training data and the code under an MIT licence. Where TypeSafe sells Jev as a general decision engine, Cua bet on a single narrow specialist and gave it away.
Both companies borrow the same vocabulary, calling these "System One" models: fast intuition rather than slow reasoning. The framing is deliberate. They are not trying to replace the large language model so much as to strip the routine, high-volume decisions out of it, the classifying and routing and form-filling that does not need a poet. Using a costly reasoning model to monitor another model's behaviour, Almeida argues, makes little sense when a cheap probability will do.
Jev is named after William Stanley Jevons, the economist who observed that when something gets cheaper we tend to use far more of it. That is the bet in one word. If intelligence becomes nearly free for narrow tasks, it will not stay in a handful of mega-apps; it will spread into everything, quietly, the way the early internet did. Whether a probability without a paragraph is what most users actually want is the open question. But for the first time in a while, the frontier is not the biggest model in the room. It is the smallest one that knows exactly what it does not know.