Qwen3.6-27B-Text-Only-W4A16-g128-MTP-ToolCalling
=================================================

This work is a modified derivative of:

    Qwen/Qwen3.6-27B
    Copyright 2026 Alibaba Cloud
    Licensed under the Apache License, Version 2.0
    https://huggingface.co/Qwen/Qwen3.6-27B

Modifications made in this derivative (Apache-2.0 Section 4(b)):

  - The multimodal vision encoder was removed; only the text backbone
    (Qwen3_5ForCausalLM) is retained and quantized. Input/output is text only.

  - Weight-only quantization with AutoRound 0.12.3: 4-bit integer, symmetric,
    group_size 128 ("W4A16"), packing auto_round:auto_gptq.

  - The following layers were kept at BF16 (not quantized):
      * lm_head, token embeddings
      * every linear-attention layer's in_proj_a / in_proj_b
      * the 15 mtp.* tensors (mtp.fc, mtp.layers.0.*, mtp.norm,
        mtp.pre_fc_norm_embedding, mtp.pre_fc_norm_hidden) — grafted
        verbatim from the BF16 upstream base after quantization, so
        vLLM's Qwen3_5MTP loader recognises them and native speculative
        decoding works out of the box.

  - Calibration corpus for the W4A16 quantization: pile-val (calibration
    pool of 512 lines; AutoRound nsamples=256).

No fine-tuning and no additional training data were applied. This is a pure
post-training quantization of the publicly released base model, with the MTP
head taken verbatim from the same public base.

This derivative is distributed under the Apache License, Version 2.0; a verbatim
copy of the license is provided in the file LICENSE.

"Qwen" and "Qwen3.6" are used solely to identify the upstream model. This
release is not official and implies no endorsement by or affiliation with the
Qwen team or Alibaba Cloud.