Qev
Open decision models fine-tuned from Qwen at 0.8B to 9B that return a choice and per-option probabilities for Choice, Noul and Score.
Overview
Qev fine-tunes Qwen into a Jev-style decision model. You give it a state, a question and candidate answers, and it returns a choice with a probability for every option; one model handles Choice, Noul (yes/no) and Score (ordered levels). Checkpoints are published on Hugging Face at 0.8B, 2B, 4B and 9B, along with the training data, and the repository includes the architecture, training and evaluation code, a Python interface and checkpoint tools. English and Chinese are supported.
The request format mirrors Jev's: state plus a questions object with type, instructions and criteria. Training examples use the same shape with a label per question, so you can fine-tune on your own data. The README compares Qev against Kev, JevAny, NanoJev and Jev as a reference on several benchmarks, and credits Jared Palmer's Kev for conventions it adapts. It targets researchers and teams who want a self-hosted decision model.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle this? | billing · shipping |
| Yes / no | Any yes/no question about the state | — |
| Score | Ordered rating questions with zero-based levels | — |
How it works
- Load a Qev checkpoint once with Qev.from_pretrained
- Submit a state and typed questions
- Read the choice and the probability for each option
- Optionally fine-tune on labelled examples in the same format
Strengths
- Open weights at several sizes for local inference
- Same request shape as Jev, so prompts carry over
- Training and evaluation code included, with a CPU smoke test
Caveats
- Needs a GPU for the larger checkpoints and a specific PyTorch build
- Benchmark comparisons are the authors' own
Getting started
python -m pip install -e . (Python 3.12, PyTorch 2.8.0), run python scripts/smoke.py --out runs/smoke, then Qev.from_pretrained("AustinFu/Qev-9B").At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
- Has tests
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Qwen fine-tuned into decision models at 2B, 4B and 9B, with the training data and evaluation code. English and Chinese.
MrJev/awesome-jev
Frequently asked questions
What does Qev decide?
Any Choice, Noul or Score question you pose about a state, returning probabilities for every option.
Does it need a TypeSafe key?
No; it is an open model you run yourself from Hugging Face weights.
Can I train it on my own data?
Yes; training examples use the inference format plus a label per question.
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