Int4 weight-only quantization of Qwen/Qwen3.8-27B, in compressed-tensors format for vLLM. 19.42 GB, down from 55.56 GB — it fits a 24 GB card with room for a useful context window.

Unofficial and unaffiliated with Alibaba/Qwen. All model capabilities, evaluations and limitations belong to the original model card — see the base model for those.

What was changed

Weights were quantized from bfloat16 to int4, group size 128, symmetric, weight-only (activations stay 16-bit) using llmcompressor.model_free_ptq. No calibration data was used and the model was never loaded — the quantizer operates directly on the safetensors. Architecture, tokenizer, chat template and processor configs are the vendor's, unmodified.

component	precision	size
language-model linears (64 layers)	int4 g128	12.56 GB (64.7%)
`embed_tokens` + `lm_head` (untied)	bfloat16	5.09 GB (26.2%)
vision tower (`model.visual`)	bfloat16	0.92 GB (4.7%)
MTP speculator head (`mtp.*`)	bfloat16	0.85 GB (4.4%)
conv1d kernels, norms, biases	bfloat16	0.01 GB
total		19.42 GB

Four things are deliberately left at 16-bit:

model.visual.* — vLLM builds multimodal towers with quant_config=None, so a checkpoint carrying quantized vision weights cannot be loaded.
mtp.* — the built-in multi-token-prediction speculator head, loaded through vLLM's speculative-decoding path rather than the main stack.
linear_attn.conv1d — 3-D causal-convolution kernels in the gated-delta-net blocks, shape (10240, 1, 4). Not Linear layers, and quantizers reject them outright.
lm_head + embed_tokens — precision-sensitive, and lm_head is untied here.

The linear-attention projections (in_proj_*, out_proj) are quantized; only the convolution kernels beside them are excluded.

Usage

Runs on released vLLM — the architecture has been supported since 0.25.1, so no nightly build is required:

bash

vllm serve GotoAI-Inc/Qwen3.8-27B-W4A16 \
  --max-model-len 65536 \
  --enable-auto-tool-choice --tool-call-parser qwen3_xml \
  --reasoning-parser qwen3

Do not pass --quantization; compressed-tensors is detected from config.json. The int4 W4A16 scheme uses Marlin kernels and runs on compute capability 7.5 and above.

Controlling reasoning depth

The chat template defaults to reasoning_effort='xhigh', which produces long deliberation. Both knobs below are template variables, passed through chat_template_kwargs:

json

{"chat_template_kwargs": {"reasoning_effort": "low"}}   // xhigh (default) | medium | low
{"chat_template_kwargs": {"enable_thinking": false}}    // skip thinking entirely

Set a server-wide default with --default-chat-template-kwargs '{"reasoning_effort": "low"}'; request-level values still win. preserve_thinking: false drops earlier turns' thinking from history, which matters for long multi-turn sessions.

Context

262144 tokens natively. The base model card documents a YaRN recipe for 1M tokens via --hf-overrides plus VLLM_ALLOW_LONG_MAX_MODEL_LEN=1; that is not configured here, and RoPE scaling costs quality at short contexts, so enable it only if you need it.

Reproducing this checkpoint

Built with llm-quantizer:

./llmq.py run --profile qwen3.8-27b

which is equivalent to:

python

# llmcompressor==0.13.1a20260814, compressed-tensors==0.18.1a20260818,
# transformers==5.15.1, torch==2.13.0
from llmcompressor import model_free_ptq

model_free_ptq(
    model_stub="Qwen/Qwen3.8-27B",
    save_directory="Qwen3.8-27B-W4A16",
    scheme="W4A16",
    ignore=["re:.*visual.*", "re:.*mtp.*", "re:.*\\.conv1d$",
            "lm_head", "re:.*embed_tokens.*"],
    device="cuda:0",
)

The source ships as 18 shards of ~4 GB, and a job holds one shard at a time, so the build peaks at a few GB of VRAM — no re-sharding needed and no large GPU required.

Evaluation

No benchmarks have been run. Data-free round-to-nearest quantization degrades quality more than a calibrated (GPTQ/AWQ) or QAT build; how much, for your task, is unmeasured here. Treat the published Qwen3.8 numbers as describing the bfloat16 model, not this one.

For an agentic model the informative checks are well-formed reasoning_content and clean multi-step tool calls rather than perplexity: structured emission degrades before fluency does.

License

Apache 2.0, inherited from the base model — the vendor's LICENSE is included unmodified. "Qwen" is Alibaba's mark; this repository is not endorsed by or affiliated with Alibaba.