Laya (official package)

The open-source decision engine Jev's ideas are compared against: one forward pass per typed question, three checkpoints, and a router that picks the right one per request.

33 msone typed question on a T4 (single forward pass)

7.2 msper question when batched

3checkpoints: Laya 421M · multilingual 322M · typed-decisions 421M

100+languages covered by the multilingual checkpoint

0generated tokens — answers are typed values

Open source Python · 100+ languages github.com/NandhaKishorM/laya

What the Laya package is

Laya is an open-source, non-autoregressive System 1 decision engine. You hand it a state (an email, a ticket, a JSON document) and a set of typed questions, and it answers them in a single forward pass — choice from a list, score in a range, noul for a calibrated yes/no. Because nothing is generated, there is no prose to parse and no output-token bill.

Published figures: 33 ms for one question on a T4, 7.2 ms per question batched. That is the same economic argument as Jev's, implemented in the open and runnable on your own hardware.

Three checkpoints, one router

Checkpoint	Encoder	Params	Context	Use it for
`laya`	ModernBERT-large	421M	512	English
`laya-multilingual`	mmBERT-base	322M	1024	100+ languages, about 2× faster
`laya-typed-decisions`	ModernBERT-large	421M	1024	the typed-decision workflows

The built-in Router is the recommended entry point: it detects script and language in sub-milliseconds and dispatches to the best checkpoint, so a pipeline that receives mixed-language traffic does not need its own language detection step. Preloading keeps routing under roughly 35 ms.

Run it

pip install laya

import laya
from laya import Router

router = Router(preload=True)

state = {
    "from": "[email protected]",
    "subject": "Duplicate charge on invoice #4411",
    "body": "Hi, we were billed twice for March. Please refund the duplicate today.",
}

questions = {
    "department": {"type": "choice",
                   "instructions": "Who should handle this?",
                   "criteria": ["billing", "technical", "sales"]},
}

Python 3.10+; the floor comes from huggingface_hub 1.x, transformers 5.x and torch 2.14. Weights ship on Hugging Face (convaiinnovations/laya, laya-multilingual), with a hosted Space demo and a Colab notebook for a first run.

How it is trained, and why that matters

Laya is trained with reinforcement learning against strictly proper scoring rules — the project calls it RLCD. Proper scoring rules are the technical reason a reported probability can be trusted: a model that is paid for honest probabilities beats one that is paid for confident answers. That is the same design goal Jev states for its confidence values, which is why the two are worth comparing at all.

Practically, this is the checkpoint to reach for when you want Jev-style typed decisions without a hosted dependency: same question shape, open weights, and a licence (Apache-2.0) that allows commercial use with attribution.

Source

Repository: NandhaKishorM/laya · package pip install laya · models on Hugging Face under convaiinnovations · Apache-2.0.

Other repositories in this series

Laya-MLX — Apple Silicon runtime
Laya vs Jev (T-Rex arena) — Head-to-head benchmark

Keep browsing: all open-source repos · Jev cases · what Jev is · the Laya guide

Last updated: 2026-09-23 · sources & corrections · every card links to its author's original post