license: apache-2.0
base_model:
- orcarouter/Qwen3.8-27B-Uncensored
tags: - qwen
- qwen3.8
- mxfp4
- quantized
- uncensored
- abliterated
- vllm
- quark
language: - en
library_name: transformers
pipeline_tag: text-generation
Qwen3.8-27B-Uncensored-MXFP4-awq
MXFP4 quantization of orcarouter/Qwen3.8-27B-Uncensored, made with AMD Quark and exported as an AWQ-style checkpoint for serving with vLLM on AMD RDNA4 GPUs (gfx1201 / RX 9000 series) with ROCm.
Model Details
| Source model | orcarouter/Qwen3.8-27B-Uncensored |
| Architecture | Qwen3.5-27B (Qwen3_5ForConditionalGeneration) |
| Quantization tool | AMD Quark |
| Quantization | MXFP4 (Quark export, quant_method: quark) |
| KV cache | Post-RoPE KV quantization enabled |
| Context length | 262,144 |
| Hidden size / layers | 5120 / 64 |
| Attention | 24 heads, 4 KV heads (GQA) |
| Vocab size | 248,320 |
| Multimodal | Vision tower present (excluded from quantization) |
Quantization excludes lm_head and the vision tower, so those run at higher precision.
Usage
Serve with vLLM:
vllm serve just1moremodel/Qwen3.8-27B-Uncensored-MXFP4-awq \
--port 8081
Requires a vLLM build with Quark/MXFP4 support (ROCm on RDNA4 recommended).
Intended Use
- Uncensored/abliterated fine-tune — intended for research, creative writing, and local inference where refusal behavior is not desired.
- Single-file
model.safetensors(~19 GB), sized to fit consumer GPUs with 24 GB+ VRAM (depending on context length and offloading).
Limitations
- MXFP4 is an aggressive quantization; expect some quality loss vs bf16/fp8.
- Uncensored models may produce harmful or objectionable output. You are responsible for how you use this model.
- Not tested for production or safety-critical use.
License
Apache-2.0, inherited from the base model.