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

WarpQuant Qwen3.8-27B R16E4H4 Text

This is the text-only Qwen3.8-27B checkpoint produced with WarpQuant. It contains the causal language model, tokenizer, and chat template. The vision tower and MTP layer are not included.

Technical report · Code · Multimodal model

Format

WarpQuant applies a signed Hadamard rotation and 3-bit group quantization to 400 projection tensors. An Output-Fisher score selects a 0.05 bpw weak-column recovery budget in the original activation coordinates. The token embedding and output head use grouped INT4.

The first release checkpoint stores reconstructed tensors in BF16 for direct Transformers evaluation. Its physical safetensor size is therefore larger than the analytical packed payload.

The text-only export was compared against the source conditional-generation checkpoint on the same tokenized prompt. The last-token logits matched exactly (max_abs_logit_error = 0, mean_abs_logit_error = 0, identical top-1), followed by a deterministic generation smoke test.

Evaluation

Commonsense is the macro average of fixed 1,000-example HellaSwag, WinoGrande, and PIQA screens. GSM8K uses the same first 500 examples, 5-shot prompts, and flexible-extract accuracy for all four models.

Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "HarimxChoi/WarpQuant-Qwen3.8-27B-R16E4H4-Text"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.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}
}