Modelpublic

HarimxChoi/WarpQuant-Qwen3.5-4B-R16E4-Text

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes47downloads

Model Card

WarpQuant Qwen3.5-4B R16E4 Text

Text-only export of the Qwen3.5-4B WarpQuant checkpoint. Projection weights use a signed-Hadamard block-GPTQ INT3 base, Output-Fisher selects the BF16 recovery columns, and the token embedding uses group-128 INT4.

Technical report · GitHub · VLM model

Payload and evaluation

The denominator is the 4,205,751,296-parameter text model; vision tensors are excluded.

Format	Text bpw	Payload	WikiText-2 PPL ↓	ARC-299 ↑	MMLU-13,943 ↑
BF16	16.00	7.846 GiB	8.3885	45.82	39.58
Q4KM	5.13	2.523 GiB	8.5472	48.83	39.48
IQ3_M	4.09	2.015 GiB	10.6976	42.81	37.41
WarpQuant Fisher R16E4	3.6514	1.788 GiB	9.2494	46.15	38.13

The repository stores the quantized values in BF16-compatible safetensors. The reported payload is the packed-equivalent analytical size including codes, scales, recovery values, and column indices.

Usage

Qwen3.5 currently requires the latest Transformers main branch:

bash

pip install "transformers @ git+https://github.com/huggingface/transformers.git@main"

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "HarimxChoi/WarpQuant-Qwen3.5-4B-R16E4-Text"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)

Citation

bibtex

@misc{choi2026warpquant,
  author = {Harim Choi},
  title = {WarpQuant: Dual-Domain LLM Quantization via Hadamard Rotation and Output-Fisher Sensitivity},
  year = {2026},
  url = {https://harimxchoi.github.io/projects/warpquant/}
}