license: apache-2.0 base_model: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive base_model_relation: quantized library_name: transformers pipeline_tag: image-text-to-text model_type: qwen3_5_moe tags:

qwen3.6
qwen3_5_moe
nvfp4
compressed-tensors
quantized
vllm
blackwell
rtx-5090
sm120
moe
multimodal
agentic
tool-calling
coding
uncensored
conversational

Qwen3.6 35B A3B HauhauCS Uncensored NVFP4

Uncensored Qwen3.6 35B A3B MoE quantized to NVFP4 compressed-tensors for vLLM on NVIDIA Blackwell / RTX 5090.

35B total / 3B active MoE
HauhauCS Aggressive uncensored source
Conservative NVFP4 profile: linear attention and MTP kept in bf16 for quality
NVFP4 W4A4 compressed-tensors
~22 GB
Runs on one RTX 5090
100K-131K text context target
vLLM native loading

The model files are placed at the repository root so Hugging Face shows the weights in the right-side download panel and vllm serve can load the repo directly. The repo intentionally keeps a single root weight set to avoid full-repo snapshot downloads pulling multiple profile variants.

Download

hf download lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 \
  --local-dir ./qwen36-35b-a3b-hauhaucs-nvfp4

vLLM quickstart

VLLM_NVFP4_GEMM_BACKEND=marlin \
vllm serve lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 \
  --served-model-name qwen36-35b-a3b-hauhaucs-nvfp4 \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --max-model-len 131072 \
  --max-num-seqs 1 \
  --max-num-batched-tokens 4096 \
  --gpu-memory-utilization 0.90 \
  --enable-prefix-caching \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --trust-remote-code

Local path quickstart:

hf download lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 \
  --local-dir ./qwen36-35b-a3b-hauhaucs-nvfp4

VLLM_NVFP4_GEMM_BACKEND=marlin \
vllm serve ./qwen36-35b-a3b-hauhaucs-nvfp4 \
  --served-model-name qwen36-35b-a3b-hauhaucs-nvfp4 \
  --quantization compressed-tensors \
  --kv-cache-dtype fp8 \
  --max-model-len 131072 \
  --max-num-seqs 1 \
  --max-num-batched-tokens 4096 \
  --gpu-memory-utilization 0.90 \
  --enable-prefix-caching \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --reasoning-parser qwen3 \
  --trust-remote-code

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 x 1024 tokens
MTP tensors copied from Qwen/Qwen3.6-35B-A3B
Converted using li-yifei/gguf-to-nvfp4

Pipeline:

Q8_K_P GGUF -> step1_convert_qwen36_moe.py -> HF bf16 -> step2_quantize_qwen36_moe.py -> NVFP4

Source models

Uncensored source: HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive
Original base: Qwen/Qwen3.6-35B-A3B

Acknowledgments

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

Qwen3.6 35B A3B HauhauCS Uncensored NVFP4

Uncensored Qwen3.6 35B A3B MoE quantized to NVFP4 compressed-tensors for vLLM on NVIDIA Blackwell / RTX 5090.

35B total / 3B active MoE

HauhauCS Aggressive uncensored source

Conservative NVFP4 profile: linear attention and MTP kept in bf16 for quality

NVFP4 W4A4 compressed-tensors

~22 GB

Runs on one RTX 5090

100K-131K text context target

vLLM native loading

vLLM quickstart

VLLM_NVFP4_GEMM_BACKEND=marlin \ vllm serve lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 \ --served-model-name qwen36-35b-a3b-hauhaucs-nvfp4 \ --quantization compressed-tensors \ --kv-cache-dtype fp8 \ --max-model-len 131072 \ --max-num-seqs 1 \ --max-num-batched-tokens 4096 \ --gpu-memory-utilization 0.90 \ --enable-prefix-caching \ --enable-auto-tool-choice \ --tool-call-parser qwen3_coder \ --reasoning-parser qwen3 \ --trust-remote-code

Local path quickstart:

hf download lyf/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive-NVFP4 \ --local-dir ./qwen36-35b-a3b-hauhaucs-nvfp4 VLLM_NVFP4_GEMM_BACKEND=marlin \ vllm serve ./qwen36-35b-a3b-hauhaucs-nvfp4 \ --served-model-name qwen36-35b-a3b-hauhaucs-nvfp4 \ --quantization compressed-tensors \ --kv-cache-dtype fp8 \ --max-model-len 131072 \ --max-num-seqs 1 \ --max-num-batched-tokens 4096 \ --gpu-memory-utilization 0.90 \ --enable-prefix-caching \ --enable-auto-tool-choice \ --tool-call-parser qwen3_coder \ --reasoning-parser qwen3 \ --trust-remote-code

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 x 1024 tokens

MTP tensors copied from Qwen/Qwen3.6-35B-A3B

Converted using li-yifei/gguf-to-nvfp4

Pipeline:

Q8_K_P GGUF -> step1_convert_qwen36_moe.py -> HF bf16 -> step2_quantize_qwen36_moe.py -> NVFP4