Is this another Jev under a different name?

No, and the README compares against it directly. It reports Jev assigning zero probability to the true label on 16% of DAIR Emotion examples and scoring 0.425 on 77-label Banking77. Laya is an independent Convai Innovations project, not affiliated with TypeSafe.

Are the shipped checkpoints actually usable zero-shot?

Near chance, on the project's own numbers. Base checkpoints score 0.36 and 0.35 against a 0.318 baseline, and raw calibration is weak. The 0.766 figure belongs to a fine-tuned typed-decisions checkpoint measured over 2,000 decisions, not to zero-shot use.

Will it read my whole document or quietly cut it off?

It ships with a 1,024-token limit that truncates long documents, so you must pass max_len=8192 and name the checkpoint multilingual, or long mostly-English text routes to the English model. Even then, results degrade past roughly 4,000 tokens.

Do I need a GPU, and is there a TypeScript or .NET path?

A GPU is what makes inference fast; Python 3.10+ and CPU-only builds are documented. A TypeScript and Node port ships as laya-ts, and the latest release adds a C# inference port, but that port has no NuGet package yet and lacks long prediction, hooks and batching.

What are Laya's three checkpoints and their context lengths?

laya (ModernBERT-large, 421M, 512 tokens, English), laya-multilingual (mmBERT-base, 322M, 1024 tokens up to 8192, 100+ languages) and laya-typed-decisions (ModernBERT-large, 421M, 1024 tokens).

How fast is Laya and how is that measured?

33 ms for one question and 7.2 ms per question batched, measured on a T4 in a single forward pass, with no text generation to parse.

How much does fine-tuning improve Laya's accuracy?

On the typed-decisions benchmark of 2,000 decisions across four workflows, the fine-tuned laya-typed-decisions checkpoint scores 0.766 accuracy against 0.362 for the base English checkpoint.

Jev directory

CommunityJev Repos & SDKs

Laya

Open multilingual System 1 decision models with published checkpoints for choice, score and noul questions, plus a router that dispatches each request to the right checkpoint in one forward pass.

See highlights Open github.com

Stars: 33.3k
Language: Python
License: APACHE-2.0
Reading time: 3 min
Published by: Community
Author: NandhaKishorM
Last verified: Sep 2026

NandhaKishorM/laya

User repository on GitHub

View on GitHub

Non-autoregressive System 1 decision engine. Typed choice, score and yes/no decisions over any text in a single forward pass, in 100+ languages, with a router that picks the right checkpoint per request.

Forks: 3k
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Multilingual, non-autoregressive System 1 decision engine. Typed decisions over 100+ languages in a single forward pass — 33 ms — trained with reinforcement learning against strictly proper scoring rules (RLCD), with a router that picks the right checkpoint per request. With uv, run uv add laya in a uv project or uv pip install laya in a virtual environment.

Quoted from the project's README. Read the full README.

Tagscommunity python open-models routing classification

Works forJev for Routing Jev for Scoring

Used inJev for Software & DevTools

Highlights

Three checkpoints cover English (ModernBERT-large, 421M) and 100+ languages (mmBERT-base, 322M), with a Router picking per request.
Answers typed choice, score and noul questions in one forward pass: 33 ms for one question and 7.2 ms per question batched on a T4.
Trained with RLCD, reinforcement learning against strictly proper scoring rules, with a predict_shortlist path for many labels.
The README reports Jev gave zero probability to the true label on 16% of DAIR Emotion examples and scored 0.425 on 77-label Banking77.
Base checkpoints score near chance zero-shot (0.36 and 0.35 against a 0.318 baseline) while the fine-tuned checkpoint reaches 0.766.

Quickstart

python

import laya
from laya import Router

router = Router(preload=True)
state = {"body": "We were billed twice. Please refund the duplicate today."}
questions = {"department": {"type": "choice", "instructions": "Which department should handle this?", "criteria": {"billing": "invoices, payments, refunds", "technical": "bugs, outages", "sales": "pricing"}}}
res = router.predict(state, questions)
print(res["answers"]["department"]["choice"])

Watch out

Apache-2.0, developed by Convai Innovations. Needs Python 3.10+ and a GPU for fast inference; Jev comparisons use different prompts and sample sizes, raw calibration is weak, and it is not affiliated with TypeSafe.

Questions about this entry

Short answers drawn from the linked project or page.

Reactions & coverage

Posts, threads, and videos about this entry from around the web.

Hacker News: ASTGL replay chose local Laya routing over hosted Jev on one Mac Studio setup.

Jmeg8r15 pointsDoes an open-weight decision model beat a hosted one? Jev vs. LayaRead the thread on Hacker News

Reddit: High-traffic thread on open Laya weights versus TypeSafe's hosted Jev launch.

Reddit: A walkthrough of routing between models with Jev.

X: Standard One SH (StandardThinking)

From the Jev guides

Original write-ups on what this entry is for.

The Open Reimplementations: Local Models That Speak System One

Sep 22, 2026

Closed weights have not stopped the ecosystem from rebuilding the shape in the open. API-compatible servers, retrained heads on small models, and inference-time tricks that skip generation entirely.

Jev Repos & SDKs#open-models#research#local

Read article

From the Jev AI model community

Posts from builders shipping with Jev right now.

Follow @typesafeai