LitJev

September 18, 2026 · View on GitHub

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

LitJev reproduces the decision layer of Jev 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, then:

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 defined in the Jev documentation: the same POST /v1/systemone path, the same model / state / questions request body, the same choice, score and noul question types, and the same answers and usage response. Code written for one can talk to the other by changing the base URL.

Send a request with curl (the server must be running):

curl --fail-with-body http://127.0.0.1:8000/v1/systemone \
  -H 'Content-Type: application/json' \
  --data-binary @examples/request.json

The state is what the model looks at; each entry in questions is one decision about that state:

{
  "model": "litjev",
  "state": "Customer writes: my order arrived broken and I need it replaced today.",
  "questions": {
    "intent": {"type": "choice", "instructions": "What does the customer want?",
               "criteria": {"refund": "Money back", "replace": "A replacement",
                            "info": "Just information"}},
    "urgency": {"type": "score", "instructions": "How urgent is this?",
                "criteria": ["Not urgent", "Urgent", "Critical"]},
    "escalate": {"type": "noul", "instructions": "Should a human take over?"}
  }
}

The response returns one typed answer per question ID:

{
  "model": "Qwen/Qwen3.8-27B",
  "answers": {
    "intent": {"type": "choice", "choice": "replace",
               "probabilities": {"refund": 0.12, "replace": 0.85, "info": 0.03},
               "confidence": 0.71},
    "urgency": {"type": "score", "score": 1.6,
                "legend": {"0": "Not urgent", "1": "Urgent", "2": "Critical"},
                "probabilities": {"0": 0.05, "1": 0.30, "2": 0.65}, "confidence": 0.42},
    "escalate": {"type": "noul", "noul": 0.78}
  },
  "usage": {"input_tokens": 96, "output_tokens": 0}
}

choice returns the winning option key, score the probability-weighted level index, and noul the probability of yes. Values above are illustrative. Up to ten questions fit in one request. From Python:

import requests

response = requests.post("http://127.0.0.1:8000/v1/systemone", json=request_body)
answers = response.json()["answers"]
if answers["escalate"]["noul"] > 0.5:
    hand_off_to_human()

Switching between LitJev and the Jev API

Because the schema is identical, the same request body works against the hosted Jev service. Only three things differ:

	LitJev (local)	Jev (hosted)
URL	`http://127.0.0.1:8000/v1/systemone`	`https://api.typesafe.ai/v1/systemone`
Auth header	none	`Authorization: Bearer <API_KEY>`
`model`	`litjev` or the loaded checkpoint ID	`jev-latest`

Point your client at the other URL, set the header and model name, and everything else stays the same. Prototype locally on your own GPU, then switch to Jev, or the other way round. The numeric probabilities and confidence values will differ between the two, since LitJev runs a different model; only the contract is shared.

Modules

Decision API — POST /v1/systemone, described above. Interactive docs at /docs, full reference in HTTP API.
Playground — browser UI at / with probability bars, logits and timings.
Benchmarking — MMLU-Pro direct-answer scoring plus Doom and chess played from screenshots, with a /film replay viewer. See benchmarking.
Calibration — optional post-hoc temperature fitting. See how it works.

@software{litjev2026,
  author  = {{ZhengxuYu}},
  title   = {LitJev: A Jev Decision Layer for Off-the-Shelf LLMs},
  year    = {2026},
  version = {0.1.0},
  url     = {https://github.com/zhengxuyu/litjev}
}

LitJev

Supported models

Quick start

Using the API

Switching between LitJev and the Jev API

Modules

Documentation

License

Citation