Laya
Multilingual non-autoregressive decision engine answering choice, score and yes/no questions in one ~33 ms pass, with a checkpoint router.
Overview
Laya, from Convai Innovations, is an open, multilingual, non-autoregressive System 1 decision engine. It answers typed choice, score and yes/no questions over text in more than 100 languages in a single forward pass of about 33 ms, trained with reinforcement learning against strictly proper scoring rules (RLCD). A router detects the script of each request and dispatches it to the right checkpoint among three, including a multilingual model that reads up to 8,192 tokens when max_len is raised from its default of 1,024.
It installs from PyPI with optional extras for an HTTP server, an MCP server, LangChain and LangGraph, LlamaIndex, CrewAI, ONNX Runtime, a GPU fast path and pydantic structured outputs, and has a TypeScript package for Node and the browser. Weights are on Hugging Face with a demo Space and a Colab notebook. The README publishes long-document accuracy (16 to 18 of 20 correct up to about 4,000 tokens, more variable beyond) and, per a listing, ends its benchmarks with a limits section naming datasets it does not generalise to. Many other projects use Laya as a local backend.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option fits the text? | — |
| Score | Where does the text fall on the scale? | — |
| Yes / no | Is this statement true of the text? | — |
How it works
- pip install laya and load the router
- Send text and typed questions
- Router picks a checkpoint by script and model choice
- Model returns calibrated probabilities in one pass
- Optionally serve over HTTP or MCP
Strengths
- Open weights and local inference
- 100+ languages
- Many framework integrations
- Honest limits and long-context measurements
Caveats
- Default 1,024-token limit truncates long documents
- Accuracy varies beyond about 4,000 tokens
- Does not generalise to some datasets, per its limits section
Getting started
python -m pip install layaAt a glance
- Type
- Library
- Status
- beta
- Language
- Python
- Runtime
- PyTorch library, HTTP and MCP servers; TypeScript package
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
- Stars
- 32,782
Use cases and integrations
How curated lists describe it
Open alternative: non-autoregressive decision model that answers `choice`, `score`, and `noul` questions with RLCD-trained calibrated probabilities in a single ~35 ms forward pass, published on PyPI and Hugging Face.
yibie/awesome-jevConvai Innovations' multilingual non-autoregressive decision model: Choice, Score, and Noul in one forward pass, with published weights, a PyPI package, and a router that picks a checkpoint per request
AnotiaWang/awesome-decision-modelsNon-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.
wh000wh000/awesome-jev-liveNon-autoregressive decision engine over 100+ languages: three checkpoints and a router that detects the script and dispatches per request. Its benchmarks end with a limits section naming the datasets it does not generalise to and the headline figure that came from a training split.
MrJev/awesome-jev
Frequently asked questions
How fast is Laya?
About 33 ms per decision in a single forward pass, per the README.
Which languages does it support?
More than 100, with a router choosing the checkpoint by script.
Can it read long documents?
The multilingual model reads up to 8,192 tokens with max_len=8192; accuracy varies past about 4,000.
Related projects
Open-source, self-hosted System One decision engine (NotJev : Kev) with a Jev-style API on top of Ollama or any OpenAI-compatible model.
LibrarybetaTypeScriptMangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++