local-jev
Browser-only Jev-style decisions in Chinese or English using a zero-shot NLI model on ONNX Runtime Web, with honest limits.
Overview
This local Jev (zhuanqianfish), documented in Chinese, runs Jev-shaped decisions entirely in the browser with ONNX Runtime Web and transformers.js: state and typed questions in, structured judgments out, with zero token cost and no data leaving the machine. It uses a zero-shot NLI classification model, by default a multilingual mDeBERTa model that handles Chinese, with WebGPU when available and WASM otherwise. The first load downloads about 280 MB, then uses the browser cache.
Questions follow Jev's types: noul takes a declarative statement and optional negation, choice takes options with descriptions, and score takes ordered levels and returns an expected level. The README gives practical advice: phrase noul questions as statements because NLI models expect premise and hypothesis, describe each choice clearly, and keep score levels evenly spaced. Its honest limitations section notes that probabilities are not calibrated like Jev's RLCD training, questions run one at a time rather than in parallel, long text is truncated around 512 tokens, and English NLI models are more accurate on English tasks.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | The customer is asking for a refund or billing correction (statement) | — |
| Choice | Which team should handle this? | billing · tech · sales |
| Score | How angry is the customer? | not angry at all · a bit dissatisfied · very angry · extremely angry |
How it works
- Load the NLI model in the browser
- Call jev.ask(state, questions)
- Each question becomes NLI hypotheses scored against the state
- Return Jev-shaped answers with probabilities
Strengths
- Runs fully in the browser
- Zero token cost
- Chinese support via a multilingual model
- Clear guidance on question phrasing
Caveats
- Not calibrated like Jev
- Questions evaluated sequentially
- About 512-token context
Getting started
npm install && npm run dev, then open http://localhost:5173 in Chrome or Edge and load the model.At a glance
- Type
- Project
- Status
- prototype
- Language
- JavaScript
- Runtime
- browser (ONNX Runtime Web, WebGPU/WASM)
- Has code
- yes
Use cases and integrations
How curated lists describe it
本地版 Jev：浏览器端 System One 判断模型（noul/choice/score），纯本地 ONNX 推理，零 token 费用.
wh000wh000/awesome-jev-live
Frequently asked questions
How does it work without Jev?
It uses a zero-shot NLI model in the browser to score statements and options.
Are its probabilities calibrated?
No, the README says they are softmax outputs and thresholds must be tuned.
Does it support Chinese?
Yes, via a multilingual mDeBERTa model.
Related projects
Mangaba 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++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython