vLLM-oriented quantization of ProCreations/grug-v1.1-qwen-3.8-27b, pinned to revision 3ab073b4bb06dc8a83e819a33485469f32b945ba. Activation-aware W4A16 asymmetric INT4 with group size 128, calibrated on 128 Ultrachat samples at 1024 tokens. The vision tower, embeddings, LM head, GatedDeltaNet a/b gates, and MTP namespace remain BF16.

The trained MTP head is not included in this repository. Use the corresponding -mtp-... repository for speculative decoding.

Serve

bash

vllm serve ProCreations/grug-v1.1-qwen-3.8-27b-awq-int4 --max-model-len 32768 \
  --reasoning-parser qwen3 --enable-auto-tool-choice --tool-call-parser qwen3_coder

The exact build script and structural report are included under reproduce/ and quantization_report.json.