Is this Jev and TypeSafe or something else?

Laya. The digest credits the Convai Innovations Laya model and package with a community ONNX export; it shares the typed-decision shape of choice, score, and noul in one forward pass with no tokens, but it is a separate local-serving stack.

ONNX or torch on the Pi, which backend should I point at?

Point at ONNX. The digest reports ONNX fp32 roughly 1.3 to 1.65 times faster than torch and numerically matching it, while int8 visibly degrades the noul heads, so the default model file is fp32.

What are the Pi-specific setup catches?

Stock PyPI aarch64 torch is a CUDA build that dies with SIGILL on Cortex-A72, so the image uses the CPU wheel; intra-op threads are limited with full graph optimization, and fair Wi-Fi is isolated via a script.

What can I try without reading code first?

Open the Snake stream, the five-example showcase covering email triage, moderation, routing, triage, and multilingual calls, plus the FAQ page for scores and explanations, all from any LAN machine.

What license does laya-pi use, and what hardware does it target?

Apache License 2.0, with bundled fonts under SIL OFL 1.1. It targets a Raspberry Pi 4 (Cortex-A72) and runs entirely on CPU with no cloud, packaging the stack as a single Docker image with four roles.

How do I start it, and what is downloaded on first boot?

Run docker compose up -d --build. On first start the model weights are downloaded from Hugging Face into the laya-models volume, which takes a few minutes; then open http:// :8080/ for the Snake stream and http:// :8091/ to try the decision API.

Does the Pi run inference when nobody is watching?

No. The Snake plays only while at least one spectator is connected, so with no audience the Pi does no inference. An exhibition mode script can isolate the Wi-Fi hotspot from the LAN, and typical latency is about 2.2 s per decision on ONNX with four threads versus about 4.2 s on torch.

CommunityJev Models & Reimplementations

Laya Pi

Self-contained Raspberry Pi 4 stack running Laya System-1 typed decisions fully on CPU: one Docker image serving ONNX and torch backends behind the same `/v1/systemone` protocol.

See highlights

Open github.com

Stars: 1
Language: Python
License: APACHE-2.0
Published by: Community
Author: trinacria-tech
Last verified: Oct 2026

trinacria-tech/laya-pi

User repository on GitHub

View on GitHub

una API systemone sul tuo raspberry pi4+ 8gb ram

Forks: 0
Last pushed: Updated Oct 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Esegui Laya — decisioni tipizzate "System-1" in un solo passaggio — interamente su un Raspberry Pi 4, solo CPU, senza cloud. Lo stack è un'unica immagine Docker con quattro ruoli: Tutti i servizi parlano lo stesso protocollo POST /v1/systemone + GET /health, quindi qualsiasi client Laya può puntare a entrambi i backend senza modifiche. Al primo avvio i pesi del modello vengono scaricati da Hugging Face nel volume laya-models: servono alcuni minuti. Poi apri, da qualsiasi macchina della LAN:

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

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Serves `/v1/systemone` plus `/health` from `laya-onnx` on 8001, showcase on 8091, snake on 8080, optional torch on 8000.
Downloads model weights from Hugging Face into the `laya-models` volume on first boot over several minutes.
Plays Snake in easy, hard, and RL modes streamed over SSE, staying idle with no spectators to spare inference.
Uses CPU torch wheels, fp32 ONNX, capped single-checkpoint memory, and limited intra-op threads on the Pi 4.

Watch out

First boot downloads weights and takes minutes; two torch checkpoints leave under 400 MB free, so one stays resident. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

gut

Python helper that turns judgment calls into one-liners like likely, classify and rate, defaulting to Jev while also running on local or OpenAI-compatible backends.