# Wang Xiaolong (CountingSheep): vev

> Jev-like decision model that also reads images: a LoRA fine-tune of Qwen3.5-4B that answers yes/no, multiple-choice and graded questions about text, JSON or screenshots with a probability per allowed answer, behind a Jev-compatible API. Non-commercial licence.

- Page: https://systemonemodels.ai/countingsheep/vev
- API: https://api.systemonemodels.ai/v1/models/countingsheep/vev
- Download: `pip install systemonemodels && systemone pull countingsheep/vev`

## Facts

| | |
|---|---|
| Maker | Wang Xiaolong (CountingSheep) (https://systemonemodels.ai/countingsheep) |
| Decides | choice, score, noul |
| Architecture | vev |
| Base model | qwen/qwen3.5-4b |
| Parameters | 4.0B |
| Context | 32K tokens |
| Licence | cc-by-nc-4.0 |
| Availability | Open weights |
| Released | 2026-09-30 |
| Latest version | 2026.09 |

## Reported evaluation

Suite: JevBench public subset (231 items), the maker's own run with his evals harness, v0.1.0. Numbers are the publisher's own.

- Decision accuracy: 76.6%

## Model card

<!-- generated by scripts/seed_catalog.py; edit content/models/catalog.yaml -->

# vev-4b

Jev-like decision model that also reads images: a LoRA fine-tune of Qwen3.5-4B that answers yes/no, multiple-choice and graded questions about text, JSON or screenshots with a probability per allowed answer, behind a Jev-compatible API. Non-commercial licence.

A v0.1 research preview by Wang Xiaolong. Answers come from next-token probabilities over the answer tokens, with no classifier head. It matches Jev's request format but is not affiliated with TypeSafe. Probabilities are not calibrated (the maker's ECE ranges 0.01 to 0.14), and reversing the option order flips the top answer on 13 to 17% of questions. The maker's own comparison puts Jev ahead on JevBench, nimble and Kev's transfer set. Sibling vev-9b (0.823 on JevBench's public subset in the maker's run) and LoRA-only repos. The weights are CC BY-NC 4.0 because some training data is research-only; the code is Apache-2.0.

## What it decides

- **choice** — picks one option from a set
- **score** — places the input on an ordered scale
- **noul** — answers a yes/no question with one probability

## At a glance

| | |
|---|---|
| Parameters | 4B |
| Base model | `Qwen/Qwen3.5-4B` |
| Maker | Wang Xiaolong (CountingSheep) |
| Released | 2026-09-30 |
| License | cc-by-nc-4.0 |
| Reported accuracy | 76.6% |

## Get the weights

```bash
pip install systemonemodels
systemone pull countingsheep/vev
```

The files are served from the maker's Hugging Face repository, [`CountingSheep/vev-4b`](https://huggingface.co/CountingSheep/vev-4b), and verified against the checksums recorded here.

## Read more

- [Model card](https://huggingface.co/CountingSheep/vev-4b)
- [vev-9b](https://huggingface.co/CountingSheep/vev-9b)
- [Code](https://github.com/Xiaooolong/vev)

---

*This page was opened by System One for Wang Xiaolong (CountingSheep), who can claim the organisation and take it over at any time.*

---

From System One Models — https://systemonemodels.ai/ · every System One model: https://systemonemodels.ai/system-one-models