Models that pick from a list

Use it for quick choices. It can sort tickets and emails. It can find what a user wants in a chat. It can pick a tool from a fixed list. It can check a reply for personal data, as a guardrail. It can also score answers against a checklist, as in evals. The LLM still does the writing and the hard thinking.

Models that sort things into groups are called classifiers. They are not new. Teams have trained small ones for years, one per job. The new idea is one general model, built only for typed answers. You describe the choice in plain words, and call it like an LLM.

The psychologist Daniel Kahneman described two ways of thinking, in his 2011 book Thinking, Fast and Slow. System 1 is fast and automatic, like seeing that a face is angry. System 2 is slow and careful, like working out 17 × 24 on paper.

An LLM is more like System 2. A reasoning model even writes many thinking tokens before its answer. A decision model is built to be System 1. It is non-autoregressive: it does not write step by step. So almost nothing is written, or billed as output.

This kind of model got a lot of attention in September 2026. As of October 2026, two early examples are Jev and Laya.

• Jev, by TypeSafe AI, is closed and in early access. You call it over the internet, through its maker's API (a way for programs to call it) or some AI gateways. So your text leaves your own systems.
• Laya, by Convai Innovations, has open weights under Apache-2.0. You can download it and run it yourself. Then your team must run and secure it. It reads text but never writes it.

Laya's model card lists limits. On its own, the base model scores close to guessing on its maker's typed test. It does far better after extra training for the task. It gets weaker with more than about 20 options. The English version reads only 512 tokens per question.

This field is new. Treat most numbers as claims from the makers. TypeSafe says Jev is up to about 200 times faster than LLMs, on tasks it tested.

One early study put Jev into a real system. It is a September 2026 preprint, so other experts have not checked it yet. In its updated version, the typical wait for a decision fell by about 23 to 65 percent. That is compared with the fastest LLM it tested. The cost per correct decision fell by about 60 to 80 percent. Jev got slightly fewer answers exactly right. So the gain was real, but much smaller than the claim. The network and the rest of the system still take time.

Also, an LLM can be forced to answer only from a list. This is called structured output. Then a decision model's edge is speed and cost, not a valid form. Is it right enough for your question? Test that on your own cases.