LVJ / field notes Updated 2026-09-22
Decision models / evidence-led comparison

laya vs jev

As of September 2026, Laya and Jev are decision-model products for software that needs typed answers instead of generated prose. Laya provides open weights that can be self-hosted under Apache 2.0. Jev is a hosted TypeSafe AI service with a published input price and no public model weights. Neither product is a general-purpose chat model.

Watch / both

TypeSafe AI Jev vs. Laya: what they are and the controversy so far

Adam Gardner (@agardnerit) · both

Open this video on YouTube
01

What the two models are

Laya

Laya is a family of non-autoregressive System 1 decision models from Convai Innovations. Its main English checkpoint uses a 421M-parameter ModernBERT-large encoder. It accepts text, email, ticket, or JSON state and typed questions, then returns structured answers in one forward pass.

Jev

Jev is TypeSafe AI's first System One model. Software sends unstructured state and predefined questions, then receives typed answers and probabilities. It is trained with Reinforcement Learning for Calibrated Decisions and uses parallel sampling.

02

Comparison table

DimensionLayaJev
Primary role Open-weight structured decision model Hosted structured decision model
Output Typed answers and probabilities Typed answers and probabilities
Main public checkpoint 421M parameters not published
Other checkpoints 322M multilingual; 421M typed-decisions Versioned hosted model, including Jev 1.13.0 in published tests
Sampling approach Non-autoregressive, single forward pass Parallel sampling
License or access Apache 2.0 weights; self-hostable Hosted TypeSafe API; early access in reviewed sources
Language coverage Model family described as covering 100+ languages; multilingual card displays 51 languages not published as a comparable benchmark
Published price No hosted Laya API price; local inference has hardware and operations costs $0.042 per million input tokens; output listed as free
Latency evidence 32.8ms for one routed question in the official T4 measurement Independent published p50 measurements range from 236ms to 276ms and include hosted API latency
Maintenance Convai Innovations open-source repositories and model cards TypeSafe AI hosted service
Watch / Laya

Laya vs Jev: Laya Github - Tutorial Open-Source AI Routing With Calibrated Probabilities

Alex Hitt (@alexander-hitt) · Laya

Open this video on YouTube
03

The benchmark has two different stories

Laya's official typed-decisions model card reports 0.766 for its fine-tuned checkpoint and 0.727 for the published Jev 1.13.0 result. The card also states that the Jev figures were published elsewhere and that the Laya result belongs to a checkpoint fine-tuned on that benchmark's training split.

A single Reddit head-to-head post reports the opposite direction: 92.9% for Jev, 65.3% for Laya's main 421M checkpoint, and 71.1% for Laya typed-decisions. The limits matter: it was one post, the test data was generated by Fable 5.1, and the post was only a few hours old when reviewed, before community voting had settled.

04

Context and data boundaries

The main English Laya checkpoint is documented with a 512-token context setting. In the Reddit test, Jev reportedly found a buried fact in 450 of 450 notes with inputs reaching about 4,000 tokens, while Laya degraded when the relevant fact moved beyond its available window.

This was still a single early community test using AI-generated data, not a universal context benchmark. It does show why context size, data location, and network path belong in any serious comparison.

05

Which one should you choose?

L

Choose Laya when

You need weights on your own machines, offline inference, Apache 2.0 licensing, or control over fine-tuning and calibration.

J

Choose Jev when

You prefer a managed API, do not want to operate model weights, or need a hosted decision service around predefined schemas.

06

FAQ

Is Laya really more accurate than Jev?

There is no unconditional answer in the public evidence. Laya's official typed-decisions result is higher on that benchmark, while a separate Reddit test using synthetic Fable-generated data reports a large Jev advantage.

Source · Source

Is Jev open source?

The reviewed sources do not show public Jev weights or a self-hosting option. Jev is presented as a TypeSafe AI hosted service.

Source · Source

Can I run Laya locally?

Yes. The Laya model card provides Python installation and loading examples, and the weights are listed under Apache 2.0.

Source

Can I run Jev locally?

No local Jev deployment is documented in the reviewed sources. The public integration path is a hosted TypeSafe API.

Source · Source

Why do the Laya and Jev benchmark results disagree?

The tests use different data, labels, prompts, checkpoints, context limits, and deployment environments. They are evidence about different test setups, not two measurements of one identical experiment.

Source · Source

07

Sources and scope

  • Laya model definition, checkpoints, usage, and license no date
    Open source official claim
  • Laya multilingual checkpoint and language metadata no date
    Open source official claim
  • Laya typed-decisions benchmark and checkpoint no date
    Open source official claim
  • Jev definition, RLCD, pricing, and early-access status 2026-09-15
    Open source official claim
  • Jev availability, openness, and claim limitations 2026-09-18
    Open source independent analysis
  • Jev integration and structured question types 2026-09-17
    Open source independent analysis
  • Jev hosted latency and benchmark conditions no date
    Open source independent measurement
  • Jev p50 latency range and decision benchmark no date
    Open source independent measurement
  • Reddit head-to-head synthetic benchmark 2026-09-22
    Open source community test
  • Hacker News discussion about prior work and architecture similarity 2026-09-20
    Open source community feedback