license: apache-2.0
base_model:

• Qwen/Qwen3.8-27B
  base_model_relation: quantized
  pipeline_tag: image-text-to-text
  library_name: transformers
  language:
• en
• zh
  tags:
• qwen
• qwen3
• qwen3.8
• uncensored
• abliterated
• vision-language
• function-calling
• compressed-tensors
• w4a16
• int4
• vllm
• rtx3090

Qwen3.8-27B-Uncensored — W4A16 (RTX 3090 build)

A 4-bit weight-only (W4A16) quantization of the uncensored / abliterated
Qwen3.8-27B vision-language model,
sized to run on a single 24 GB GPU (e.g. RTX 3090).

This is the exact quant used to serve an agentic assistant ("Sage") live on an
RTX 3090 in day-to-day use, so it's known to load and run under vLLM on
consumer 24 GB hardware — not just a theoretical export.

What it is

• Base model: Qwen3.8-27B (VL, function-calling, reasoning), architecture
  Qwen3_5ForConditionalGeneration.
• Uncensoring: the refusal-removed ("abliterated") build published by
  OrcaRouter under Apache-2.0. This repo is a
  quantization of that build.
• Quantization: 4-bit weight-only, produced with AutoRound and exported in
  the compressed-tensors pack-quantized format:
  • num_bits: 4, type: int, symmetric: true, group_size: 128,
    strategy: group
  • Vision-tower linears are kept at higher precision (listed under ignore),
    so image understanding is preserved.
• Footprint: ~15 GB of weights → fits comfortably in 24 GB with room for
  KV cache at a useful context length.

Serving with vLLM

vllm serve <path-or-repo>/Qwen3.8-27B-Uncensored-W4A16 \
  --host 0.0.0.0 --port 8000 \
  --served-model-name qwen3.8-27b \
  --gpu-memory-utilization 0.90 \
  --max-model-len 32768 \
  --enable-auto-tool-choice --tool-call-parser hermes

Then hit the OpenAI-compatible endpoint at http://localhost:8000/v1. Adjust
--max-model-len to trade context length against KV-cache memory on a 24 GB
card. A recent vLLM with compressed-tensors support is required.

Provenance & attribution

• Qwen — the Qwen3.8-27B base model (© the Qwen team).
• OrcaRouter — the Apache-2.0 abliterated build this repo quantizes.
• This repo — 4-bit W4A16 (compressed-tensors) quantization for 24 GB GPUs.

License

Apache-2.0, inherited from the upstream abliterated build. See the bundled
LICENSE. You must retain the license and attribution when redistributing.

Responsible use

This is an uncensored / abliterated model: its built-in refusal behaviour
has been removed, so it will attempt requests an aligned model would decline.
You are responsible for how you deploy it, for any guardrails you add, and for
compliance with the laws and platform rules that apply to you.