What Is a Decision Model?

An LLM is autoregressive: it predicts the next token, appends it, and repeats until it has produced an answer, one token at a time. A decision model skips that loop. You give it a state and a typed question (pick one of these options, rate this on a scale, answer this yes/no), and it returns an answer with a probability attached in one pass, the same way a classifier does, not the way a chat model replies.

The two products publicly framing this as a distinct model category as of September 2026 are TypeSafe AI's Jev, described as a 'System 1' model, and Convai Innovations' Laya, an open-weight model released shortly after and positioned as an alternative to it. Both target the same job: the decision, routing, and classification steps inside an AI workflow that don't need free-form text, steps a general-purpose LLM has traditionally been asked to do anyway because it was already in the pipeline.

The size difference is real but not the point of the design. TypeSafe has not published Jev's parameter count or architecture. Laya, the open-weight alternative, has: its checkpoints are 322 to 421 million parameters, built on ModernBERT and mmBERT, encoder-only architectures in the BERT family. That is small next to a modern LLM, which typically runs from several billion to over a hundred billion parameters, but the size gap is a side effect of the task, not the design goal. An encoder that classifies a fixed set of options doesn't need the capacity a model generating open-ended text does.

We loaded Laya's English checkpoint (convaiinnovations/laya on Hugging Face) and ran it locally: two questions, a sentiment check and a four-way email classification, both correct, both under 200ms on CPU alone. The model's own published numbers claim about 33ms on GPU with its fast inference path, which we did not test. Either way, decision-model latency and LLM latency are not the same kind of number: an LLM's latency scales with how many tokens it generates, a decision model's latency is closer to fixed per question.

The limitation is scope. A decision model answers the questions you defined ahead of time, in the shape you defined (a fixed list of choices, an ordinal score, a boolean). It cannot take an open-ended prompt and improvise a novel answer the way an LLM can. If the question wasn't anticipated in the schema, the model has nothing to say about it.

That makes it a component, not a replacement. The pattern both projects converge on is splitting an agent's workflow: an LLM still handles open-ended generation, drafting, summarizing, conversing, while a decision model takes over the routing and classification steps in between, where the LLM's flexibility was never needed and its latency and cost were being paid for nothing.