LitJev

# LitJev

**A reproduction of Jev: turn any Qwen model into a fast decision model.**

LitJev reproduces the decision layer of [Jev](https://docs.typesafe.ai/introduction)
on top of off-the-shelf Hugging Face checkpoints. Define questions and options, load a
Qwen model, and get typed choices with probability distributions from a single API
call. No training, no generated answer text: scores are read directly from the
model's output head.

**LitJev 是对 Jev 的复现：把 Qwen 全系列模型变成像 Jev 一样的快速决策模型。**
定义问题和选项，加载模型，即可通过 API 或浏览器前端获取选择结果与概率分布，
无需训练，也无需生成回答文本。

> Independent research project, not affiliated with or endorsed by TypeSafe AI.
> This is a hypothesis-based reproduction from public information, not the official
> Jev implementation or a connection to its API. Request/response JSON follows the
> public Jev schema; internals, confidence values and performance are not identical.
> Probabilities are not calibrated by default.

## Supported models

The full Qwen family is supported (Qwen3.x text and vision checkpoints, any size).
The default and most-tested checkpoint is `Qwen/Qwen3.8-27B` on one H100 80 GB.
Vision checkpoints are required for screenshot decisions. Other model families are
not guaranteed to work.

## Quick start

Install [uv](https://docs.astral.sh/uv/getting-started/installation/), then:

```bash
git clone https://github.com/zhengxuyu/litjev.git
cd litjev
uv run --locked litjev --model Qwen/Qwen3.8-27B
```

Open **http://127.0.0.1:8000/**, load the example, and submit. The first request
downloads and loads the model, which can take several minutes. Use
`--model /path/to/checkpoint` for a local checkpoint and `--device-map cuda:0` to pin
a GPU. Not yet published to PyPI; the commands above run this checkout.

## Using the API

LitJev uses **exactly the schema de