🚀 Qwen36-27B-GPTQ-Pro-4Bit

Overview

Qwen3.6-27B-GPTQ-Pro-4bit 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.

At a glance

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 (14 visible artifacts total)

Quick start

vLLM (documented configuration)

vllm serve groxaxo/Qwen3.6-27B-GPTQ-Pro-4bit \
  --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.

Welcome to Qwen36-27B-GPTQ-Pro-4Bit – a titan of reasoning and generation, elegantly squeezed into a remarkably efficient 4-bit package. It punches leagues above its weight class while keeping your VRAM happy and your inference speeds blazingly fast! Thank you Qwen team for another amazing model.

🌟 Why the "Pro"?

This isn't your average quantization. We used the GPTQ-Pro framework combined with the FOEM (First-Order Error Metric) approach. This advanced technique carefully preserves the most critical weights during the 4-bit compression process by evaluating the exact impact of quantization on the model's loss landscape.

The result?

• Near-Lossless Performance: Enjoy the profound reasoning, coding prowess, and vast knowledge of a 27 Billion parameter model, but with a drastically reduced memory footprint.
• Marlin Optimized: Ready out-of-the-box for Marlin kernels to deliver maximum token-per-second throughput in serving engines like vLLM.
• Consumer Hardware Friendly: Fit a massive 27B powerhouse model on consumer GPUs with room to spare for massive context lengths!

This repository contains a 4-bit GPTQ-Pro quantization of unsloth/Qwen3.6-27B, produced with GPTQModel and the FOEM/GPTAQ-style quality settings used in the GPTQ-Pro project.

Source project: https://github.com/groxaxo/GPTQ-Pro

Deployment

vLLM

CUDA_VISIBLE_DEVICES=0,1 vllm serve groxaxo/Qwen3.6-27B-GPTQ-Pro-4Bit \
  --dtype float16 \
  --quantization gptq_marlin \
  --disable-custom-all-reduce \
  --tensor-parallel-size 2 \
  --max-model-len 132144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --gpu-memory-utilization 0.92

Local path

CUDA_VISIBLE_DEVICES=0,1 vllm serve /path/to/Qwen3.6-27B-GPTQ-Pro-4Bit \
  --dtype float16 \
  --quantization gptq_marlin \
  --disable-custom-all-reduce \
  --tensor-parallel-size 2 \
  --max-model-len 132144

Transformers

from gptqmodel import BACKEND, GPTQModel

model = GPTQModel.load(
    "groxaxo/Qwen3.6-27B-GPTQ-Pro-4Bit",
    backend=BACKEND.GPTQ_MARLIN,
    device="cuda:0",
)

print(model.generate("Write a short deployment checklist.", max_new_tokens=64)[0])

Notes

• Tested with tensor parallel size 2 on RTX 3090 GPUs.
• Use float16 and gptq_marlin for the most reliable vLLM startup path.
• The quantization and serving workflow lives in the GPTQ-Pro repository above.
• MTP/speculative decoding is detected by vLLM for this model, but on 2x RTX 3090 the exact --max-model-len 262144 launch OOMs during KV-cache setup.
• The working local vLLM configuration I verified is --max-model-len 65536 with --enforce-eager; that starts and serves, but the current metrics showed spec_decode_num_accepted_tokens_total=0, so it does not improve speed yet.
• If you test MTP, use --speculative-config '{"method":"mtp","num_speculative_tokens":2}' and disable thinking in the request payload when you want a plain answer.

⚡ Speed Benchmarks

Tested on 2× NVIDIA RTX 3090 with vLLM (gptq_marlin, tensor-parallel=2, float16).

Test 1: Short Prompt → 256 Tokens (Streaming)

Test 2: Medium Prompt → 512 Tokens (Non-Streaming)

Test 3: Short Burst → 64 Tokens (Streaming)

📊 Quality Evaluation

• Wikitext-2 test perplexity: 6.366 (n_ctx=1024)