HarimxChoi/WarpQuant-Qwen3.8-27B-R16E4H4

WarpQuant Qwen3.8-27B R16E4H4

This is the complete Qwen3.8-27B multimodal checkpoint produced with WarpQuant. It retains the vision tower and multimodal components while applying WarpQuant to the 64-layer text backbone.

Technical report · Code · Text-only model

Format

The text backbone uses signed Hadamard rotation, 3-bit group quantization, block-GPTQ reconstruction, and Output-Fisher weak-column recovery. Token embeddings and the language-model head use group-128 INT4.

Text-backbone evaluation

GSM8K uses the same first 500 examples, 5-shot prompts, and flexible-extract accuracy for all four models.

KV cache and activation ablation

Use

from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "HarimxChoi/WarpQuant-Qwen3.8-27B-R16E4H4"
processor = AutoProcessor.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(model_id, device_map="auto")

Citation

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