Run Laya on your Mac

Apple Silicon is the best-covered platform in the Laya ecosystem. Independent ports run the upstream checkpoints on the MLX GPU, through Core ML on the Neural Engine, or inside native Swift apps. Each port publishes its own speed and fidelity measurements, all taken on the author's own Mac.

308 projects

SnakeGame_Laya

Okbatti

Uses a fine-tuned Laya model to choose moves in a Snake game

2·GitHub repo

PythonMLX

laya-mlx-zh

ZLHAOOO

Provides Chinese fine-tuned Laya weights, training data, and evaluations for Apple Silicon MLX

2·GitHub repo

MLXPython

design-os-generative-ui

jangtrinh

Builds a generative UI engine using local Laya-MLX decisions and a TypeSafe JEV cascade router

2·GitHub repo

Node.jsMLX

pacman-arena

hakan-gecili

Compares Laya with other local decision models playing Pac-Man under identical conditions

2·GitHub repo

PythonMLX

s1-tui

tnaftali

Provides a terminal interface for typed decisions using local Laya or the hosted Jev API

2·GitHub repo

PythonMLX

laya-modernbert-decision-90pct

ameerhmz5

Fine-tunes Laya into a ModernBERT decision model and offers deployment formats

473·HF model

PythonMLX

Laya-Multilingual-MXFP8

sahilchachra

An MXFP8 MLX quantization of Laya multilingual's encoder backbone for Apple Silicon

274·HF model

MLXRuntimes & ports

laya-toolguard-mlx

a3harva

Fine-tunes Laya in MLX to assess coding-agent tool calls for safety risks

110·HF model

MLXPython

laya-coreml-adaptive-fp16

aac6fef

Provides an adaptive FP16 Core ML conversion of the Laya model for Apple silicon

40·HF model

Core MLRuntimes & ports

Laya-Multilingual-CoreAI

mlboydaisuke

Exports Laya multilingual as an on-device Apple model for macOS and iPhone

13·HF model

Core MLRuntimes & ports

laya-multilingual-arc

goktugoguz

Fine-tunes Laya to classify Arc USDC transfers and answer questions about them

10·HF model

PythonMLX

fin-laya-mlx-6-rounds

ya9919

A 0.42B-parameter financial decision model trained on CHFS, CGSS, CFPS and CLDS data

10·HF model

MLXNVIDIA / CUDA

laya-multilingual-stellar

goktugoguz

Fine-tunes Laya to classify Stellar operations and answer questions about them

5·HF model

PythonMLX

laya-mobile

charioteer

Converts the Laya model to Core ML and ONNX for on-device use on iOS and Android

5·HF model

Core MLONNX

laya-en-fp32-swev

danielamitay

Exports Laya as a Core ML model for local typed decisions through the Swev Swift package

1·HF model

Core MLSwift

laya-typed-decisions-421m-coreml-ane

NMTZ

Provides fixed-shape Core ML artifacts converted from the 421M Laya typed-decisions checkpoint

1·HF model

Core MLRuntimes & ports

laya-multilingual-coreai

smdesai

Provides a Core AI conversion of multilingual Laya for typed decisions on Apple devices

1·HF model

Core MLSwift

laya-mlx

millat

Converts Laya to a native MLX FP16 checkpoint for inference on Apple silicon

1·HF model

MLXPython

LayaVoxprompt

MainzelMennchen

Fine-tunes multilingual Laya to classify dictation as an assistant prompt or ordinary text

1·HF model

PythonMLX

laya-snake-mlx

madhavbiplov

Fine-tunes a multilingual Laya model for calibrated snake-game decisions on Apple silicon

1·HF model

MLXFine-tunes & weights

fin-laya-mlx-3-rounds

ya9919

A 0.42B-parameter financial decision model trained on CHFS and CGSS data

1·HF model

MLXNVIDIA / CUDA

pastewhat

mizorewww

A native macOS clipboard companion recommends entries using a local Laya model or optional Jev backend

1·GitHub repo

SwiftPython

jev-codex-workbench

qualixar

Adds Jev decisions and optional local Laya-MLX routing to a Codex plugin and MCP server

1·GitHub repo

PythonMLX

pi-laya-router

AdelysAlberto

Routes Pi coding-agent prompts using local Laya classification on Apple Silicon

1·GitHub repo

SwiftRouting

Pick a runtime

MLX (Python). laya-mlx is the most widely used port. It needs Apple Silicon, Python 3.11+ and macOS 14+:

pip install laya-mlx

import laya_mlx as laya

agent = laya.load("aac6fef/laya-mlx")
result = agent.predict(
    "I was billed twice. Please refund the duplicate.",
    {"department": {"type": "choice",
                    "instructions": "Who should handle this?",
                    "criteria": ["billing", "technical", "sales"]}},
)
print(result["answers"]["department"])

Its README reports 13.42 ms P50 for one short question with the 421M English checkpoint and 7.39 ms with the 322M multilingual one, FP16 on an M3 Max.

Core ML and the Neural Engine. laya-coreml needs macOS 15+ and Python 3.11 to 3.13 (pip install laya-coreml). Its README reports 4.98 ms P50 for one short multilingual decision on the M3 Max Neural Engine. laya-apple runs the MLX GPU and the Neural Engine together and serves a Jev-compatible API:

pip install 'laya-apple[serve,ane]'
laya-apple serve                                   # http://127.0.0.1:8642

Swift. laya-mlx-swift embeds Laya in a macOS app with MLX (about 804 MiB download). LayaKit is a Swift package over the laya-coreml bundles, and FluidUse ships a LayaManager for Core ML.

PyTorch MPS. The official laya package also runs on the Mac GPU; laya-mps wraps it in a local server with low-memory modes.

Caveats

All of these are independent ports, not official Convai Innovations releases. Most publish fidelity checks against upstream answers; read them.
The Neural Engine bundle in laya-coreml has a 96-token total limit covering question, options and state. Use the 1024-token general bundle for longer input.
Latency numbers are single-machine measurements. Your Mac, input length and question count will change them.