Instructions to use matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16 with Transformers:

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

pipe = pipeline("text-generation", model="matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16")
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)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16")
model = AutoModelForMultimodalLM.from_pretrained("matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16", 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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16

SGLang

How to use matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16 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 "matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16" \
    --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": "matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

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 "matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16" \
        --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": "matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16 with Docker Model Runner:
```
docker model run hf.co/matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16
```

Qwen3.5-4B-QuantStudy-RTN-W4A16

수식과 표로 읽는 실험 보고서 / Read the report · Code and evaluation evidence

Qwen3.5-4B 전체 모델 양자화 비교에서 실제로 저장·재로딩·평가한 결과물입니다. These are the exact packed artifacts evaluated in the 2026-09-22 quantization study, not newly retuned checkpoints.

Scope and format

Base: Qwen/Qwen3.5-4B, revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.
Common targets: 200 text Linear matrices, 3,565,158,400 original weights (96 FFN, 32 full-attention, 72 gated-delta projections).
Embedding/tied output head, norms, biases, convolution/state tensors and the 48 small gated-delta in_proj_a/b matrices retain high precision. AWQ can rescale preserved tensors while retaining high precision.
모든 가중치가 4비트인 모델이 아닙니다. This custom preservation policy differs from common default community recipes bearing similar names.

Storage format: compressed-tensors pack-quantized, symmetric INT4 / group 128, BF16 activations. RTN/GPTQ/AWQ here names the algorithm, not legacy AutoGPTQ/AutoAWQ file formats. Keep quant_method=compressed-tensors; do not force a legacy GPTQ or AWQ loader. The preserved vision module is present in the weights, but this release was evaluated for text only.

Files and recipe

Candidate	Settings	Weight file	Decimal GB
R0	RTN; memoryless_minmax observer	model.safetensors	3.787

Each row uses groups of 128 input-channel weights. For G0, the 128-column GPTQ compute block is a separate setting from quantization group size. AWQ scales are folded into compatible weights; the calibration-only output-MSE search does not use final evaluation answers.

Measured quality

같은 엔진의 BF16 기준과 비교합니다. NLL은 정답 토큰에 부여한 확률의 음의 로그 평균으로, 낮을수록 좋으며 정답률과는 다릅니다.

Model	NLL ↓	PPL ↓	Δ NLL vs own BF16 ↓	KMMLU %	ARC-C norm %	IFEval strict %
B0 (vllm)	1.851833	6.3715	+0.000000	48.8	55.8	83.18
R0 (vllm)	1.940655	6.9633	+0.088822	47.2	54.8	76.52

Calibration: 128 windows × 1,024 tokens (64 Korean / 32 English / 32 code source documents); validation 32 and final 128 documents are disjoint by exact document SHA.
Final NLL: 130,944 scored tokens. KMMLU: fixed stratified 500-item subset, not the official whole benchmark. ARC-C: fixed 500 items, character-length-normalized choice scores. IFEval: 541 prompts, prompt strict, greedy, thinking off, maximum 2,048 output tokens.
Final data did not select recipe settings. Failures were not removed from denominators. Semantic duplication and pretraining exposure are not ruled out.
One calibration seed/configuration; not a claim that the algorithms were exhaustively optimized. See paired confidence intervals and failure cases before interpreting small differences.

Download and run

hf download matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16 --local-dir ./Qwen3.5-4B-QuantStudy-RTN-W4A16

For an exact immutable version, add --revision <full commit SHA> shown in the repository history. File hashes are in SHA256SUMS and ARTIFACT-MANIFEST.json.

The packed runtime was vLLM 0.26.0 + PyTorch 2.11.0 + Transformers 5.14.1 + compressed-tensors 0.17.0, on RTX 5090 / CUDA 13.0 / WSL2. Quantization used llmcompressor 0.13.0 and compressed-tensors 0.18.0.

inference_example.py uses the evaluated eager/batch-one settings. Install the pinned compatible environment, download the repository, then run:

python inference_example.py .

The example's chat prompt is illustrative; the evaluation worker is linked in the report. BF16-decompressed numerical checks are not packed INT4 speed measurements. The saved vision weights do not imply validated multimodal support for this package.

Reproducibility and attribution

See fixed environment, data and target policy, quantization-study.json, and the full report. Weight bytes and saved quantization configs are preserved from the completed experiment. Publishing adds documentation and checksums; it does not change model weights.

Base model by the Qwen team. Derived weights are distributed with the upstream Apache-2.0 LICENSE; see NOTICE for modifications. This is a learning/portfolio experiment, not an official Qwen release or a newly proposed quantization algorithm. Coding, execution automation, evaluation aggregation and documentation used Codex assistance.

Downloads last month: 138

Safetensors

Model size

5B params

Tensor type

I32

BF16

Model tree for matchachoco/Qwen3.5-4B-QuantStudy-RTN-W4A16

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Quantized

(443)

this model