Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128-preserve-vl is a GPTQ-quantized checkpoint intended for efficient GPU inference, published by groxaxo. It is intended for open-source evaluation, reproducible experimentation, and compatible local or hosted inference workflows. The wording below is deliberately limited to what can be verified from this repository's metadata and artifacts.

The repository name identifies a behavior-modified or reduced-filtering lineage. That label describes the source or conversion history; it is not a guarantee of unrestricted behavior in every prompt or runtime. Test outputs carefully before sharing or deploying them.

At a glance

Field	Details
Format	GPTQ
Source / base	`wangzhang/Qwen3.6-27B-abliterated`
Intended task	`image-text-to-text`
License	the license declared in the repository files

What is included

*.safetensors (5 files)
config.json
generation_config.json
tokenizer.json
tokenizer_config.json
processor_config.json
chat_template.jinja
quantize_config.json
Additional configuration, tokenizer, processor, or shard files (13 visible artifacts total)

Quick start

vLLM (documented configuration)

bash

vllm serve groxaxo/Qwen3.6-27B-abliterated-v2-GPTQ-Pro-FOEM-4bit-g128-preserve-vl \
  --quantization gptq_marlin \
  --dtype float16 \
  --trust-remote-code

This command is taken from the repository documentation. Adjust tensor parallelism, context length, and cache settings to match your hardware and vLLM version.

Compatibility and responsible use

Use a runtime that explicitly supports this format, architecture, and modality.
Keep configuration, tokenizer, processor, projection, and weight files from the same revision together.
Review the source model card and license before redistribution or deployment.
Hardware needs depend on parameter count, context length, cache precision, quantization, and concurrency.
Report reproducible issues with the runtime version, hardware, launch command, and a minimal example.

Quantization or conversion changes numerical behavior, memory use, and throughput relative to the source checkpoint; validate quality on your own workload.

Generated outputs may be inaccurate or unsuitable for a given use case. Users are responsible for testing behavior, applying appropriate safeguards, and complying with applicable licenses and laws.

Strict preserve-VL FOEM-enhanced GPTQ-Pro W4G128 export of Qwen3.6-27B-abliterated-v2.

Quantization recipe

Base quant recipe: GPTQ-Pro W4G128
Enhancement: FOEM + activation-weighted MSE
bits: 4
group_size: 128
desc_act: false
sym: true
true_sequential: true
act_group_aware: true
lm_head: false
FOEM: alpha=0.25, beta=0.2

Explicit preserve rules

This artifact was rerun with hard exclusion rules for fragile multimodal pieces:

visual.*
embed_tokens.*
lm_head.*
norm.*
merger.*
projector.*
patch_embed.*

Verification notes

Local post-save checks confirmed:

no quantization log hits for the preserved module patterns
visual tensors: 333
quantized tensors present
lm_head.weight remains BF16
model.language_model.embed_tokens.weight remains BF16
AutoProcessor, AutoTokenizer, and AutoConfig load successfully
architectures = ["Qwen3_5ForConditionalGeneration"]
language_model_only = false

Local launch settings used to validate the source vision path

The stable local vLLM launch on this machine was:

bash

setsid env \
  CUDA_VISIBLE_DEVICES=0,1,4 \
  CUDA_DEVICE_ORDER=PCI_BUS_ID \
  OMP_NUM_THREADS=1 \
  TOKENIZERS_PARALLELISM=false \
  NCCL_P2P_DISABLE=1 \
  NCCL_IB_DISABLE=1 \
  NCCL_NET_GDR_DISABLE=1 \
  NCCL_SHM_DISABLE=0 \
  NCCL_CUMEM_HANDLE_DISABLE=1 \
  PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True,max_split_size_mb:256 \
  /home/op/venvs/vllm-qwen36/bin/vllm serve "/home/op/models/Qwen3.6-27B-abliterated-v2" \
    --host 0.0.0.0 \
    --port 8000 \
    --tensor-parallel-size 1 \
    --pipeline-parallel-size 3 \
    --max-model-len 4096 \
    --kv-cache-dtype fp8 \
    --gpu-memory-utilization 0.98 \
    --max-num-seqs 1 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder \
    --trust-remote-code \
    --served-model-name qwen36-27b-abliterated-v2 \
    --disable-custom-all-reduce \
    --generation-config vllm \
    --enforce-eager \
    --limit-mm-per-prompt '{"image":1}'

Notes:

--max-model-len 32144 did not fit KV cache on this host.
For direct chat requests, chat_template_kwargs.enable_thinking=false was used to keep visible output in the answer channel.

Files

quantize_config.json records the explicit dynamic skip rules and FOEM metadata
processor_config.json keeps the multimodal processor config
model.safetensors.index.json and shard files contain the final export