A 4-bit weight quantization of Qwen3.5-4B (GPTQ W4A16, group size 128, compressed-tensors pack-quantized), with the lm_head untied from the embeddings and quantized to 8-bit to shrink the per-step output-projection read. Built for fast single-request (batch-1, greedy) inference and designed to pair with the DFlash speculative-decoding drafter.

Produced offline with llm-compressor GPTQModifier (4-bit symmetric int, group 128, actorder: static), calibrated on 840 public, train-split samples (ARC-Challenge, QASC, MedMCQA, MedQA-USMLE, AQuA-RAT, CommonsenseQA, Dolly). The benchmark datasets (GPQA-Diamond, MMLU-Pro, IFEval) are explicitly excluded from calibration.

This quantization is released under Apache-2.0. The underlying base model remains subject to the Qwen3.5-4B license from the Qwen team.