Décider n'est pas générer : les modèles de décision
A decision model answers with a **typed value** (a category, a score, a choice from a closed list) plus a probability, whereas a generative model produces text that then has to be read or parsed. Much of a product's AI features write nothing: they classify a ticket, pick a path, score an answer, decide whether a human should step in. For those cases the distinction changes three things: 1. **Latency**: a decision fits in the time of a click, a generation makes people wait. 2. **Cost**: you pay for a short, bounded output, not a paragraph. 3. **Specification**: the feature is described by the list of its possible outputs, not by a prompt. An answer outside the list becomes impossible instead of unlikely. **The September 2026 signal.** TypeSafe announced Jev, its first public model, sold purely for deciding (70 to 500 ms end to end, typed output, calibrated probability), first behind a waitlist, open to everyone on 21 September 2026. The numbers are the vendor's own, and Latent Space points out it does not replace an LLM: no free text, predefined outputs required. The category itself already exists under other names (classifiers, judges, routers), and that is what is worth keeping. **A decision without reasoning cannot be read back.** An LLM that gets it wrong at least leaves a sentence to read; a decision model hands you “billing, 90%” and nothing else. If the examples it learned from cover part of your customers poorly, no single answer shows it: the bias only surfaces when you count its errors, group by group, on cases your team has already sorted ([Instagram reel DdmFYslgMIe](https://www.instagram.com/reel/DdmFYslgMIe/)). Checking a decision model goes entirely through data, never through reading. **Six days later, the waitlist no longer holds.** Laya publishes the same promise under Apache 2.0: typed output, no generation, 32.8 ms median on a T4 card, and a multilingual 322-million-parameter checkpoint that fits in a gigabyte. Its benchmark puts it ahead of Jev over 2,000 decisions (0.766 against 0.727 accuracy), and admits losing past twenty possible options. What changes for you is not the ranking, it is that the category can now be tested in one command, on your own cases, without asking anyone for access. **One figure to correct along the way.** The 'fifty times faster' circulating on social media about Laya comes from no measurement: the repository claims 7.8 times. The fifty is a batch size, fifty questions fed at once, read as a speed multiplier. That is the reflex to keep for any announcement in this family: ask on what hardware, over how many decisions, and at what batch size. The same principle applies to an agent that acts. Jev Ultrafast, the Browser Use web agent, does not screenshot the page to reason: the page hands it a numbered list of what can be clicked, and a small decision model picks a number. An action outside the page becomes impossible, not merely unlikely ([Instagram reel DduETr9iOkf](https://www.instagram.com/reel/DduETr9iOkf/)).