Compare Laya 7 min Updated Sep 26, 2026

Jev vs Laya

Jev is TypeSafe’s hosted System One model. Laya is an independent Apache-2.0 local package evaluating the same three question types. Unofficial; no partnership.

This is an unofficial field-manual note. It does not rank products or treat GitHub activity as a popularity score.

What each project is

Jev is TypeSafe’s flagship System One model. You send a state plus typed questions and receive structured answers your code can branch on. TypeSafe documents the hosted call as POST https://api.typesafe.ai/v1/systemone with model jev-latest. Jev currently accepts text (strings, JSON objects, and arrays of text); images, audio, and video are not supported.

Laya is published by Convai Innovations as an Apache-2.0, non-autoregressive decision engine. Its README describes typed choice, score, and noul questions over text, email, ticket, or JSON state in a single forward pass. Weights are listed on Hugging Face under convaiinnovations/laya (and related checkpoints). Install is documented as python -m pip install laya on Python 3.10 or newer.

A compatible HTTP surface is not the same product. Laya’s authors document a laya-serve path that speaks POST /v1/systemone; that is not a TypeSafe release and is not hosted Jev.

Documented question types

Primitive	TypeSafe Jev (official docs)	Laya (project README)
Choice	Select from a defined set; returns `choice`, `probabilities`, `confidence`. A Choice accepts up to 255 options.	`choice` questions; option labels share a fixed head-token budget rather than Jev’s 255-option cap
Score	Ordered rubric; returns `score`, `probabilities`, `confidence`	`score` questions; every level needs a description (`null` is rejected)
Noul	Yes/no as `noul` in 0–1	`noul` questions

TypeSafe evaluates mixed questions in one request, in parallel and in isolation. Laya’s Router is documented as the recommended entry point: it picks among English, multilingual, and typed-decisions checkpoints.

How you call them

	Jev	Laya
Official path	TypeSafe API / SDK, or the Playground	PyPI package `laya`; optional `laya[serve]` HTTP server
Auth	`TYPESAFE_API_KEY` as `Authorization: Bearer`	No TypeSafe key. Optional `LAYA_API_KEY` on `laya-serve`
Default bind (serve extra)	Hosted `api.typesafe.ai`	`laya-serve` binds `0.0.0.0:8000` per the README

Checkpoints Laya documents

Checkpoint	Encoder (README)	Documented use
`laya` / English	ModernBERT-large, 421M, context 512	English
`laya-multilingual`	mmBERT-base, 322M, context 1024 (up to 8,192)	100+ languages
`laya-typed-decisions`	ModernBERT-large, 421M, context 1024	typed-decisions workflows

The first Hub download needs network access to Hugging Face. Routing-only CLI use is documented as working without downloading a checkpoint.

Limits the Laya README calls out when porting a Jev client

Option budget. Options share head_max_len (192 tokens on the English checkpoint, 256 on the others), not TypeSafe’s 255-option Choice cap.
Score levels. Every level needs a description.
Confidence. Laya documents confidence on choice/score as 1 minus normalised entropy, and says a threshold carried over from Jev does not transfer.