license: apache-2.0
base_model: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
base_model_relation: quantized
tags:

• qwen3.6
• nvfp4
• compressed-tensors
• quantized
• vllm
• moe
• multimodal
• image-text-to-text
• blackwell
• rtx-5090
• sm120
• uncensored
  library_name: transformers
  pipeline_tag: image-text-to-text

Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4

NVFP4 quantized version of HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive.

Conservative profile: linear_attn (30 DeltaNet/Mamba layers) and MTP kept in bf16 for best quality. Follows AEON-7/RedHatAI approach.

Quantization Recipe

recipe = QuantizationModifier(
    targets="Linear", scheme="NVFP4",
    ignore=["lm_head", "re:.*visual.*", "re:.*mlp.gate$",
            "re:.*mlp.shared_expert_gate$", "re:.*linear_attn.*", "re:^mtp.*"],
)
oneshot(model=model, dataset=ds, recipe=recipe,
        max_seq_length=1024, num_calibration_samples=128,
        moe_calibrate_all_experts=True, pipeline="basic")

• Calibration: HuggingFaceH4/ultrachat_200k, 128 samples × 1024 tokens
• MTP tensors copied from Qwen/Qwen3.6-35B-A3B (not present in GGUF)

Deployment (vLLM)

Vision + text smoke-tested on RTX 5090

This repository has been smoke-tested locally on an RTX 5090 with vllm/vllm-openai:v0.21.0-cu130-local, compressed-tensors, NVFP4 Marlin GEMM, FP8 KV cache, and a real image chat.completions request.

VLLM_USE_FLASHINFER_MOE_FP4=0 \
VLLM_NVFP4_GEMM_BACKEND=marlin \
vllm serve ./Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 \
  --served-model-name qwen36-35b-a3b-hauhaucs-nvfp4 \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --gpu-memory-utilization 0.90 \
  --max-model-len 4096 \
  --max-num-seqs 1 \
  --max-num-batched-tokens 1024 \
  --trust-remote-code

For short non-thinking answers, pass chat_template_kwargs at the top level of the OpenAI-compatible request:

{
  "chat_template_kwargs": {"enable_thinking": false}
}

Text-only long context

VLLM_USE_FLASHINFER_MOE_FP4=0 \
VLLM_NVFP4_GEMM_BACKEND=marlin \
vllm serve ./Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --gpu-memory-utilization 0.95 \
  --max-model-len 100000 \
  --max-num-seqs 1 \
  --reasoning-parser qwen3 \
  --language-model-only \
  --trust-remote-code

Pipeline

Converted using li-yifei/gguf-to-nvfp4:

Q8_K_P GGUF → step1_convert_qwen36_moe.py → HF bf16 → step2_quantize_qwen36_moe.py → NVFP4

Also See

• lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4-100K — Aggressive variant (linear_attn + MTP also NVFP4, smaller footprint for vision+long context)

Acknowledgments

• HauhauCS for the uncensored GGUF source
• Qwen for the base model and MTP weights
• AEON-7 and RedHatAI for conservative quantization approach reference