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-quantizedformat: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.