Jev vs. Laya: The Decision-Model War Started a Week After the Category Launched

The System One category, models that return typed decisions instead of prose, is about a week old. It already has a rival. A demo making the rounds shows Laya, an open-weights decision model, playing Tetris against cloud-hosted Jev and making moves roughly eleven times faster, running locally on a 16GB MacBook Air.

That is a real result worth understanding. It is also a narrow one, and the gap between those two statements is where the useful thinking lives.

What Laya actually is

Laya comes from ConvAI Innovations and is released under Apache 2.0, open weights you can download and run yourself. It is a 421-million-parameter model built on a ModernBERT backbone, which is a fraction of the size of a frontier text model. It answers typed questions against a state and returns calibrated probabilities, in roughly 33 milliseconds, through a Python package on ordinary consumer hardware.

Jev, from TypeSafe AI, is the model that defined the category: typed Choice, Score, and Noul questions answered against a state, with confidence attached, hosted in the cloud. Laya targets exactly that shape and removes the hosting.

Why the Tetris demo is impressive and incomplete

Tetris is a near-perfect showcase for a decision model. The option set is small and fixed, the state is fully visible, and there is no ambiguity about what a good move looks like. Latency is the binding constraint, so a local model with no network round trip wins on the dimension the game measures.

Most business decisions are not Tetris. Is this draft on brand? Should this ticket escalate? Does this claim have evidence behind it? Those have fuzzy option sets, incomplete state, and, crucially, a right answer that matters more than a fast one. A model that is eleven times faster and slightly less calibrated is a worse tool for that work, not a better one.

Where local genuinely wins

• Volume, classification and routing at scale, where per-call cloud pricing compounds fast.
• Privacy, client data, HR records, or anything you would rather not send to a vendor's endpoint.
• Latency-bound loops, interactive tools, game-like control, anything inside a tight feedback cycle.
• Offline and edge, workflows that cannot assume a connection.

Where hosted still wins

• Nobody has to maintain it, no model updates, no hardware, no ops rota.
• Consistency across a team, everyone hits the same version of the same model.
• Evaluation maturity, a newer open model has far less public evidence about how its confidence behaves under pressure.

What this means if you're building workflows

The honest headline is not "Laya beat Jev." It is that the decision layer is now a competitive category with both a hosted and an open option inside its first month. That is good news for anyone building workflow systems: it means typed, confidence-bearing decisions are going to get cheap, and the differentiator moves up a level, to how you define your options, where you set your thresholds, and who reviews the flagged cases.

That part has never been a model problem. It is a design problem, and it is still yours.

Practical advice: prototype on the hosted model because it is faster to reach a working decision, then move to local weights when volume, privacy, or latency make the case. Do not start local to save money on a workflow you have not proven yet.