Instructions to use quevirus/Xevn-4B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use quevirus/Xevn-4B with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("image-text-to-text", model="quevirus/Xevn-4B")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("quevirus/Xevn-4B")
model = AutoModelForMultimodalLM.from_pretrained("quevirus/Xevn-4B", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use quevirus/Xevn-4B with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "quevirus/Xevn-4B"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "quevirus/Xevn-4B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/quevirus/Xevn-4B

SGLang

How to use quevirus/Xevn-4B with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "quevirus/Xevn-4B" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "quevirus/Xevn-4B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "quevirus/Xevn-4B" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "quevirus/Xevn-4B",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use quevirus/Xevn-4B with Docker Model Runner:
```
docker model run hf.co/quevirus/Xevn-4B
```

Xevn-4B

English | 한국어

Xevn is a 4B decision model built for Korean, trained for Korean understanding, knowledge, and judgment.

Xevn follows the Jev decision format. It reads a question and its options once and returns the answer together with a probability for every option. There is no text generation, so a request costs one forward pass.

This repository provides the bf16 Transformers weights. Weight size: 9.3198 GB. It retains the base vision tensors for HF layout compatibility; the decision helper uses text only. The tables below report the MLX bf16 reference and its quantized variants. Reference MLX bf16 equal-mixture HTTP p50: 88.37 ms.

Tasks

Type	Input	Output
`choice`	A question and up to 26 options	Probability for each option
`noul`	A yes/no question with criteria	Probability of yes
`score`	An item and a graded scale	Probability for each level

Benchmarks

Korean

Accuracy (%)

Benchmark	Measures	Xevn	JevK5	Plumb	Decider	Mica	Tev1
KoBEST	Reasoning and commonsense	89.7	69.7	84.2	74.6	72.7	67.7
KorMedMCQA	Korean medical licensing exams	69.4	65.5	65.5	64.3	65.8	59.7
HAE-RAE	Korean knowledge	58.3	56.5	54.9	53.7	54.6	53.2
Belebele	Reading comprehension	89.0	87.4	86.6	87.2	86.3	87.6
KMMLU	Expert knowledge across subjects	56.3	52.8	55.1	54.1	54.4	47.5
CSAT-QA	Korean SAT reading	68.8	69.4	67.7	66.2	60.5	65.1
CLIcK	Korean culture and language	66.5	67.3	67.2	64.9	64.3	62.9
KMMLU-1035	Expert knowledge, second set	50.1	48.0	48.3	51.0	48.2	43.6

Headline = (Intelligence + Calibration) / 2, the JevBench score that combines accuracy and calibration.

Benchmark	Xevn	JevK5	Plumb	Decider	Mica	Tev1
KoBEST	92.0	83.7	84.2	82.0	85.3	87.1
KorMedMCQA	78.7	76.7	74.0	73.1	70.4	67.1
HAE-RAE	70.2	69.3	68.2	63.9	62.1	61.0
Belebele	90.3	89.7	87.7	89.3	86.4	89.2
KMMLU	68.5	63.3	67.8	58.1	58.5	55.3
CSAT-QA	76.4	76.6	77.3	72.3	65.7	74.4
CLIcK	72.5	74.8	73.1	68.5	68.1	68.6
KMMLU-1035	60.1	55.7	57.8	53.3	49.5	49.3

English

Accuracy (%)

Benchmark	Measures	Xevn	JevK5	Plumb	Decider	Mica	Tev1
MMLU	Knowledge across 57 subjects	76.5	74.8	74.9	74.5	75.4	71.6
ARC	Grade-school science	96.7	96.7	96.4	96.5	96.5	95.3
SciQ	Science exam questions	95.6	96.6	95.4	95.6	95.3	95.4
PIQA	Physical commonsense	87.3	87.2	86.1	88.7	84.3	84.8
Social IQa	Social commonsense	76.5	78.1	75.3	76.1	69.9	75.9
OBQA	Open-book science reasoning	88.7	88.7	87.7	91.0	85.0	79.3

Headline

Benchmark	Xevn	JevK5	Plumb	Decider	Mica	Tev1
MMLU	83.2	79.7	78.8	80.0	77.3	74.2
ARC	94.8	94.9	87.7	95.6	97.0	94.6
SciQ	94.5	94.2	88.7	94.1	95.5	94.7
PIQA	80.8	84.6	83.0	87.2	81.5	79.7
Social IQa	79.6	82.4	78.1	75.8	67.3	77.6
OBQA	84.9	86.1	78.6	89.6	86.9	83.2

Setup. Every model was run in 16-bit (bf16; Decider in its official FP16) with the same prompts and the same scoring, on an Apple M4 Max.

Speed

Latency of a single request over HTTP, bf16, Apple M4 Max, 3,000 requests.

Input length	p50	p95
up to 64 tokens	56 ms	64 ms
65 to 128 tokens	88 ms	103 ms
129 to 256 tokens	156 ms	188 ms

Models

Precision	Size	Speed (p50)	Download
16-bit	8.4 GB	88 ms	MLX
8-bit	4.5 GB	83 ms	MLX
6-bit	3.4 GB	82 ms	MLX
5-bit	2.9 GB	80 ms	MLX
4-bit	2.4 GB	78 ms	MLX
3-bit	1.8 GB	77 ms	MLX

16-bit is the bf16 format. Speed is the p50 over an equal mix of the three input lengths, measured in the same session as the Speed table. On every benchmark, 8-bit to 5-bit stay within 2 points of 16-bit accuracy and 4-bit within 3 points. 3-bit loses 7 to 21 points and is provided for experimentation.

PyTorch (Transformers) 16-bit weights: Xevn-4B.

By benchmark

Accuracy (%)

Benchmark	16-bit	8-bit	6-bit	5-bit	4-bit	3-bit
KoBEST	89.7	89.7	89.6	89.5	87.4	79.7
KorMedMCQA	69.4	69.1	69.6	68.5	66.4	50.8
HAE-RAE	58.3	58.3	56.5	57.5	56.3	40.2
Belebele	89.0	88.8	89.0	88.7	87.8	80.8
KMMLU	56.3	56.1	55.8	56.8	55.3	42.3
CSAT-QA	68.8	68.3	67.7	69.4	66.8	48.3
CLIcK	66.5	66.2	66.9	65.9	65.8	58.8
KMMLU-1035	50.1	49.9	50.5	50.9	50.1	41.5

Headline

Benchmark	16-bit	8-bit	6-bit	5-bit	4-bit	3-bit
KoBEST	92.0	91.9	92.4	91.4	90.8	80.6
KorMedMCQA	78.7	78.7	79.0	77.8	75.5	65.9
HAE-RAE	70.2	69.9	68.9	68.2	68.1	56.9
Belebele	90.3	90.0	89.6	89.2	89.6	82.8
KMMLU	68.5	68.7	68.1	69.7	69.1	55.8
CSAT-QA	76.4	76.4	74.7	74.5	74.5	61.8
CLIcK	72.5	72.2	72.6	73.1	74.5	68.0
KMMLU-1035	60.1	59.9	60.8	61.3	61.1	53.4

Intended use

Xevn is built for decisions: classification, multiple choice, yes/no judgments, and rating. It is not a chat or text-generation model. It is tuned for Korean; English input works but is not the focus.

License

Apache 2.0. Built on Qwen3.5-4B.

Quick start

pip install huggingface_hub torch "transformers>=5.2.0"
hf download quevirus/Xevn-4B --local-dir ./Xevn

Use the bundled xevn_infer.py decision helper with ./Xevn. It renders the decision prompt and reads option probabilities in one forward pass. temps.json contains the per-type readout temperatures. The benchmark and speed tables above describe the bf16 reference; the precision table reports each measured variant separately.

한국어

Xevn은 한국어 분류, 객관식 선택, 예/아니요 판단, 평점을 위한 4B 결정 모델입니다. 질문과 선택지를 한 번 읽고 답과 선택지별 확률을 반환합니다. 한국어 벤치마크 결과와 정밀도별 측정값은 위 표에 있습니다.

from xevn_infer import Xevn
model = Xevn("./Xevn")
# Pass a Jev state and a single Jev question dictionary.
result = model.predict(state, question)

Downloads last month: -

Safetensors

Model size

5B params

Tensor type

BF16

F32

Model tree for quevirus/Xevn-4B

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Finetuned

(1081)

this model