Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

edgejev

Converts open Jev reproductions (Laya, Kev, NanoJev, PlayJev) to INT8 ONNX for offline CPU decisions at about 15.6 ms per question.

ServicebetaApache-2.0
View on GitHub ↗

Overview

EdgeJev unifies open Jev and System One reproductions, including Laya, Kev, NanoJev and PlayJev, into one convert, quantise and deploy path: export to ONNX, quantise to INT8, and run a local service compatible with the official protocol. It runs fully offline on ordinary CPUs with no API key, no network and no data leaving the machine, and the runtime needs only onnxruntime, tokenizers and numpy, not PyTorch.

The README reports a single question at 15.6 ms on a four-core CPU using the same Laya weights, compared with about 314 ms for the hosted Jev API and 32.8 ms for the original Laya on a T4 GPU, with gains attributed to removing the network round trip, ONNX graph optimisation and INT8 integer kernels. The model shrinks from 1,290 MB to 324 MB. Three questions batched take 44.8 ms, which the author notes fits within voice-agent latency budgets. Documentation is mainly in Chinese.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich team should handle this request?billing · technical · sales
Yes / noDoes this message express urgency or time pressure?yes · no
ScoreHow dissatisfied is the customer?calm statement · emotional but reasonable · very angry

How it works

  1. edgejev build converts a backend (e.g. laya) to INT8 ONNX
  2. Load the model with edgejev.Agent
  3. Call system_one(state, questions) locally
  4. Read choice, probabilities and confidence

Strengths

  • Fully offline, no key
  • Very low CPU latency as reported
  • Small runtime footprint without torch

Caveats

  • Conversion step needs torch
  • Accuracy is that of the open reproductions, not hosted Jev
  • Chinese-first documentation

Getting started

uv tool install "edgejev[build]", then edgejev build --backend laya --out ./jev-int8.

At a glance

Type
Service
Status
beta
Language
Python
Runtime
ONNX Runtime (CPU), local server
Has code
yes
License
Apache-2.0

Use cases and integrations

How curated lists describe it

  • 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.

    hellogumbo/awesome-jev

Frequently asked questions

Does EdgeJev need the internet?

No, after conversion it runs fully offline.

Which models does it support?

Open reproductions such as Laya, Kev, NanoJev and PlayJev.

How fast is it?

The README reports 15.6 ms per question on a four-core CPU.

Related projects

  • decidealot

    Docker service exposing local Laya, Von, a CLM head or hosted Jev through one TypeSafe-compatible HTTP and MCP endpoint.

    MCP serverbetaPython
  • go-system-one

    Native Go runtime serving Gemma 4 12B as a Jev-like decision model with hand-tuned PTX kernels and no CGo or CUDA toolkit.

    ProjectbetaGo
  • gojev

    Go harness for System One models: hosted Jev, open Kev in-process via llama.cpp, and Cloudflare Clef, with typed helpers and a CLI.

    ServicebetaGo
  • hearim

    Go gateway exposing the Jev /v1/systemone contract over ordinary LLM backends by reading single-token label logprobs.

    ServicebetaGo
  • intern-decision-mlx

    Runs Intern-Decision-0.8B on Apple Silicon behind a local /v1/systemone endpoint that accepts screenshots for agent decisions.

    ServicebetaPython
  • jev

    TypeScript library for one-line local decisions with open models (kev-0.6b, kev-4b, open-jev), no API keys or cloud.

    ProjectbetaTypeScript