pipeline_tag: image-text-to-text base_model: orcarouter/Qwen3.8-27B-Uncensored license: apache-2.0 library_name: Model Optimizer tags:

• modelopt
• qwen3.8
• nvfp4
• fp8
• v100-skinny

Qwen3.8-27B-Uncensored NVFP4/FP8 for v100-skinny

1Cat-vLLM performs significantly faster than v100-skinny but converting the abliterated model weights to the same full NVFP4 format as the QUASAR model that 1Cat requires involves expensive training. 1Cat have also not published said training pipeline. With these weights, v100-skinny performs better than 1Cat-vLLM.

Mixed-precision quantization of orcarouter/Qwen3.8-27B-Uncensored, prepared for dnv2003/v100-skinny.

• MLP gate_proj, up_proj, down_proj, and lm_head: NVFP4, group size 16
• Full- and linear-attention projections: FP8
• Vision tower and MTP head: BF16
• Calibration: max-abs, 1,024 abisee/cnn_dailymail train samples, sequence length 512
• NVIDIA Model Optimizer source revision: 87c9f8cf83021957d1a1a575c90c9a4eaaf7ef0c

See quantization-audit.json and hf_quant_config.json for the complete machine-readable layout.

This model has had safety alignment substantially removed. Use it only for lawful, controlled research and add appropriate safeguards before deployment.