A typed-decision model reads a fixed set of questions and fills in each answer with a typed value and a confidence score. It never writes a sentence. Two of these models shipped six days apart in September 2026: TypeSafe's Jev on September 15, and Convai Innovations' Laya on September 19. Laya matters more to most developers, because you can download and run it yourself. Jev is closed, and reachable only through a waitlist.
This is a real change in shape, not just a new model. A normal chat model writes one word at a time and hopes a useful structure comes out of it. A typed-decision model is told the structure first and never leaves it. That single difference decides where each kind of model actually helps.
How it works
You hand a typed-decision model a state, such as an email or a JSON record. You also give it a list of questions, each with a fixed answer type. The model fills in every answer in one pass and attaches a confidence number to each one. According to Convai Innovations' project site, Laya answers one question in a median of 32.8 milliseconds on a Tesla T4, a low-cost graphics card common in cloud data centers.
The model never produces free text. It only ever returns a value that already fits a type you defined: a category from a fixed list, a number, or a yes-or-no. The Hugging Face model card describes Laya as a decision head added to a ModernBERT-large backbone, not a chat model with rules bolted on afterward. TypeSafe built Jev the same way. The company trained it with a method it calls Reinforcement Learning for Calibrated Decisions, rather than the approach used to train chat models.
Confidence is the other half of the idea, and it is where both models get more candid than most vendors. A model can be wrong and still be well-calibrated, as long as its confidence numbers track how often it is actually right. Laya's own materials report a calibration error of 0.466 before tuning. That number falls to the advertised 0.081 only after you refit it for each question type, using your own data. Read that as work you still have to do, not a property you inherit by downloading the weights.
Accuracy needs the same care. Laya scores 0.362 on its own zero-shot benchmark. Random guessing scores 0.318 on the same benchmark, and always picking the most common answer scores 0.461. Fine-tuned on that benchmark's training data, the same model reaches 0.766. The model card states the conclusion plainly: Laya is "a fast base to specialise, not a zero-shot decision engine."
| Model | Maker | How you get it | License | Reported latency |
|---|---|---|---|---|
| Jev | TypeSafe AI | Waitlist, hosted API | Closed | 70-500 ms end to end, per TypeSafe |
| Laya | Convai Innovations | Download the weights | Apache 2.0 | 32.8 ms median on a Tesla T4, per Convai |
Both models struggle the same way once the option list grows. Jev scores candidates in two independent stages for a question with up to 255 possible answers, according to TypeSafe. Laya's founder, Nandakishor Mukkunnoth, names a much lower ceiling for his own model: "Choice questions degrade with more than 20 options." A typed-decision model is built for a small, fixed menu of answers, not an open-ended one. Test that limit yourself before you rely on either model for a question with many possible answers.
What changed and when
TypeSafe introduced the idea first. Its Jev model, announced on September 15, scores candidates in two independent stages when a question has many options, up to 255 of them. It charges $0.042 per million input tokens, with output free. The catch is access. Jev sits behind a waitlist, and no outside benchmark of it exists yet.
Four days later, Laya arrived as an open-weights answer to the same idea. It holds 421 million parameters, a fraction of most language models' size, and ships under the Apache 2.0 license, which permits commercial use and modification. Anyone can download it, run it locally, and check what it actually does. That is not something a waitlisted API lets you do.
That gap in access is the real story here. A developer cannot check TypeSafe's calibration claims without an invitation. A developer can check Laya's claims this afternoon, on their own hardware, against their own data.
What this means for developers
Treat a typed-decision model as a replacement for one call in a pipeline, not for your whole model. The shape fits classification, routing, ranking, and tool selection: any step where an existing model already returns one of a small set of options. It does not fit anything that has to produce prose, because these models cannot write prose at all.
Read the zero-shot number, not the fine-tuned one, before you commit to either model. The gap between Laya's 0.362 and 0.766 is entirely the training split of one benchmark. If you cannot label a few thousand examples of your own decision, the zero-shot figure is the one that predicts what you will actually get.
Budget time for calibration work before trusting a confidence number from either model. Plan to refit temperature on your own data. Then log confidence against outcomes for a week, and check whether 0.7 really means seventy percent. Laya's open weights make that check something you can run yourself today. Jev's waitlist means you are trusting TypeSafe's numbers until you get access.
Check the license before you build past a prototype. Apache 2.0 lets you ship a product built on Laya. Jev's terms belong to TypeSafe and are reached only through the waitlist, so read them closely before you plan a launch around either model.
Narrow, binary decisions are where a typed-decision model already looks strong without any fine-tuning. Laya's own numbers put email spam filtering at 0.993 and phishing detection at 0.980, both far above its general zero-shot score. If your use case looks like one of those, a small typed-decision model may already beat a much larger chat model on cost and speed. If it looks like open-ended judgment instead, plan on the fine-tuning step from the start.
This article was first published on Tech AI Wire.
Also available in
Deutsch · 日本語 · Français · Español · Português
Related on Tech AI Wire
Sources
- Laya - 33ms Multilingual System 1 Decision Engine with Calibrated Probabilities - Convai Innovations
- convaiinnovations/laya model card - Hugging Face
- Introducing System One Models & Jev - TypeSafe AI
Top comments (0)