NVFP4 quantized version of Li101/Qwen3.5-35B-A3B-Uncensored-Aggressive-safetensors, which is the "Aggressive" uncensoring variant of Qwen/Qwen3.5-35B-A3B. Includes full vision encoder weights (BF16, unquantized).

Quantization Details

Detail	Value
Method	NVFP4 (compressed-tensors)
Model size	~23 GB (vs 67 GB BF16)
Language model	NVFP4 quantized
Visual encoder	BF16 (unquantized, 333 tensors, 0.89 GB)
Excluded from quantization	lm_head, MoE gates, shared expert gates, linear attention layers, visual encoder

Calibration

Samples: 512 (256 from each dataset)
Datasets:
- HuggingFaceH4/ultrachat_200k (train_sft split) — length-filtered (longest conversations selected)
- nvidia/Nemotron-Post-Training-Dataset-v2 (chat split)
Max sequence length: 1024
All experts calibrated: moe_calibrate_all_experts=True

Usage with vLLM

Please use nightly vLLM for support.

bash

vllm serve Li101/Qwen3.5-35B-A3B-Uncensored-Aggressive-NVFP4 \
    --kv-cache-dtype fp8 \
    --reasoning-parser qwen3 \
    --enable-auto-tool-choice \
    --tool-call-parser qwen3_coder

Specs

Same architecture and capabilities as Qwen/Qwen3.5-35B-A3B, including vision (image/video understanding).

Credits

Uncensoring by HauhauCS
Base model by Qwen
NVFP4 quantization and deployment by Li101