Jev & Laya FAQ

What is Jev in one sentence?

Jev is TypeSafe AI's System One model: it reads a state and a set of typed questions and returns decisions with probabilities — a choice from a list, a score in a range, or a calibrated yes/no — instead of free text.

When was Jev released and who made it?

TypeSafe AI released Jev on September 15, 2026. The team is led by Diogo Almeida, a co-creator of ChatGPT and one of the co-inventors of RLHF; the model was trained with RLCD, reinforcement learning from contrastive decisions, rather than RLHF.

What does 'System One model' mean?

It borrows the fast-versus-deliberate split from human reasoning. A System One model is built for the fast half: bounded decisions made in milliseconds, with no chain of thought to read. TypeSafe positions it as a third category alongside chat models and embeddings.

How much does Jev cost?

Community reporting puts it at roughly $0.042 per million input tokens with output tokens unmetered, because a decision barely generates output. Developers shared bills of about $2 for ~5,000 mixed decision calls. Always confirm current pricing in your own dashboard.

Which model name should I use?

Use jev-latest while experimenting, then pin jev-1.13.0 (the version it resolved to in September 2026) as soon as you tune a threshold. Floating aliases move answers under you.

Can Jev hallucinate?

It has no text to hallucinate, but it will answer the literal question you wrote. Vague instructions produce confidently wrong labels. TypeSafe's own notes also flag unreliable arithmetic, counting, hex colours and interpolation between score levels — keep those in code.

What is an open-source alternative to Jev?

Laya — a ~421M parameter non-autoregressive decision model with the same choice/score/noul answer types. It fits in 1–2 GB of RAM, runs on CPU, and has Apple MLX and Core ML ports measured at 7–14 ms and ~5 ms respectively.

Is Laya better than Jev?

On the shared public typed-decisions benchmark Laya scores 0.766 routing accuracy against 0.727 for Jev 1.13.0, and the gap is small enough that your own labelled data should decide. Laya wins decisively on privacy, offline use and marginal cost; Jev wins on operational simplicity.

Can I run these models offline?

Laya, yes: pip install laya for Python, @receptron/laya for Node via ONNX, plus MLX and Core ML ports for Apple silicon. Jev is hosted only — the API is the product.

Why do independent benchmarks of Jev look thin?

The hosted model's terms restrict public benchmarking, so most comparison numbers come from launch-week write-ups or from the open models' own scripts. That restriction is one of the reasons the open implementations attracted so much attention.

Do I still need a chat model if I use a decision model?

Yes — they do different jobs. Use a decision model for the steps that only need a label, a score or a verdict, and a chat model for the steps that must produce language. The common pattern is Jev choosing, and code or a chat model acting on the choice.

What are the answer types exactly?

choice returns one option plus a probability distribution; score returns a number in your range with per-band probabilities and a confidence; noul returns a calibrated three-way truth value (true, false, undecidable).