What really stays offline, and what needs the network?

Conversion pulls weights from Hugging Face and needs torch, but afterwards the output directory with model, tokenizer, and config files runs fully offline with no key, no telemetry, and no data leaving the machine.

How much smaller and faster is the quantized build?

The default laya path shrinks from 1290 MB to 324 MB and answers in about 15.6 ms per question on the reference CPU. Kev and other backends default to fp32 because quantization degrades their accuracy much more.

Does the converted model still agree with upstream weights?

Fp32 matches upstream laya bit-for-bit on probabilities and confidence, while kev and nanojev paths are checked against logits and argmax stability. Quantized builds trade small accuracy shifts that the digest tables break down by task.

What do I install to run versus build, and how do I switch?

Only the runtime needs onnxruntime, tokenizers, and numpy. Change the SDK base URL to the local serve port to switch from the hosted API, and build with per-backend precision flags after checking the guidance.

How do I point the official SDK at the local server?

Run edgejev serve --model ./jev-int8 --port 8009, then set TYPESAFE_BASE_URL=http://127.0.0.1:8009 and TYPESAFE_API_KEY=local. That lets you write against the official API and swap the base URL afterward.

How much does it shrink the model, and why?

From 1290 MB to 324 MB on the laya backend. The savings come from skipping the network round-trip plus a 1.7x ONNX Runtime graph optimization and a 1.2x INT8 AVX512-VNNI integer multiply-add path.

Which backends does it support, and what are their default precisions?

laya, kev, NanoJev, and PlayJev. Default precision is int8 for laya and fp32 for kev, nanojev, and playjev, because their readout-head structures absorb quantisation noise differently.

CommunityJev Demos & Experiments

Edgejev

Local offline pipeline that converts open Jev reproductions to ONNX, applies quantization, and serves typed decisions on CPU with a protocol-compatible endpoint.

See highlights

Open github.com

Stars: 21
Language: Python
Published by: Community
Author: yzfly
Last verified: Oct 2026

yzfly/edgejev

User repository on GitHub

View on GitHub

离线可用的本地类型化决策：4 核 CPU 单题 15.6ms。Local & offline Jev / System One inference on CPU — ONNX + INT8, no torch at runtime. 支持 laya / kev / PlayJev

Forks: 2
Last pushed: Updated Sep 2026

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Unifies several open reproductions into one convert, quantize, and deploy path on CPU.
Keeps inference to three runtime packages with no torch required after conversion.
Serves a System One compatible endpoint so existing SDK code switches by base URL.
Documents backend layouts, precision trade-offs, batching limits, and training helpers.

Watch out

The build step needs network access for weights and torch, and quantized accuracy varies by backend and task. 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

open-jev-typed-decision-engine

Open 150M encoder that reproduces the typed noul, choice, and score interface in one forward pass, trainable on a free GPU and runnable locally for nothing.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

intikhab49/open-jev-typed-decision-enginePython

TypeLLM

941

Library that adds schema-guaranteed typed outputs to existing autoregressive LLMs without changing weights, served through SGLang or a hosted API.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

TypeLLM/TypeLLMPython

jevify

Toolkit that makes a model behind an OpenAI-compatible endpoint or running in-process answer typed choice, score, and yes/no questions from next-token probabilities in Jev API shape.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

fidecastro/jevifyPython

jev-forge

164

End-to-end toolkit that synthesizes decision data, trains and calibrates candidate scorers, evaluates fixed splits, and serves Jev-compatible inference for interactive web decisions.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

zwliJay/jev-forgePython

stuntd

Self-hosted proxy that records typed decisions, distils them into small heads on a frozen encoder, and answers locally when confident.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

bladedevoff/stuntdPython

open-jev

Open harness that renders allowed JSON answers on a diffusion canvas, denoises for a fixed step budget, then picks the top allowed tokens from final logits with no training.

Jev Demos & Experiments#community#jevbest#catalog-import

Community

JoshuaSP/open-jevPython

From the Jev guides

Original write-ups on what this entry is for.

Per-Tick Decisions in Games

Sep 28, 2026

A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.

Jev Demos & Experiments#gaming#agents#control

Read article

Back to all resources

From the Jev community

Posts from builders shipping with Jev right now.

Follow @typesafeai