🌟 Qwen3.6-35B-A3B NVFP4 Quantization by NeuralNet 🧠🤖

This is an NVFP4-quantized version of Qwen/Qwen3.6-35B-A3B, optimized for deployment on NVIDIA Blackwell architecture GPUs using vLLM.

[!IMPORTANT] NVFP4 quantization requires NVIDIA Blackwell architecture (GB200, RTX 5000 series, etc.). This format is not compatible with Ampere, Ada Lovelace, or Hopper GPUs. If you are running on an older GPU, please use a different quantization format.

Original model: https://huggingface.co/Qwen/Qwen3.6-35B-A3B

---

Quantization Details

This model was quantized to NVFP4 (4-bit NVIDIA Floating Point) using vLLM's built-in quantization pipeline. NVFP4 leverages native FP4 Tensor Core support introduced in Blackwell GPUs, delivering significant memory savings and throughput improvements with minimal quality degradation compared to BF16.

bash
Copy code
vllm quantize \
  --model Qwen/Qwen3.6-35B-A3B \
  --quantization nvfp4 \
  --output-dir NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4

---

⚡ Deployment with vLLM

This quantized model is intended to be served using vLLM (vllm>=0.9.0 recommended).

Quick Start

bash
Copy code
vllm serve NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4 \
  --quantization nvfp4 \
  --dtype bfloat16 \
  --kv-cache-dtype fp8 \
  --max-model-len 262144 \
  --reasoning-parser qwen3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

Using a Config File

yaml
Copy code
# Deploy with: vllm serve --config config.yaml
# Optimized for NVIDIA RTX 6000 PRO (Blackwell)
# Benchmarked: ~85-90 parallel requests, up to 1000 tok/sec at higher context lengths

model: NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4
dtype: bfloat16
kv-cache-dtype: fp8
gpu-memory-utilization: 0.95
max-model-len: 262144
max-num-batched-tokens: 4096
max-num-seqs: 200
max-cudagraph-capture-size: 209
enable-prefix-caching: true
trust-remote-code: true

reasoning-parser: qwen3
enable-auto-tool-choice: true
tool-call-parser: qwen3_coder

default-chat-template-kwargs: '{"enable_thinking": false}'

download-dir: /workspace/models
host: 0.0.0.0
port: 18000

bash
Copy code
vllm serve --config config.yaml

---

💬 Chat API Usage

Qwen3.6 uses a standard chat template compatible with OpenAI-format APIs. Thinking mode is enabled by default.

Thinking Mode (Default)

python
Copy code
from openai import OpenAI

client = OpenAI(base_url="http://localhost:18000/v1", api_key="EMPTY")

messages = [{"role": "user", "content": "Your message here"}]

response = client.chat.completions.create(
    model="NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4",
    messages=messages,
    max_tokens=32768,
    temperature=1.0,
    top_p=0.95,
    extra_body={"top_k": 20},
)
print(response.choices[0].message.content)

Non-Thinking (Instruct) Mode

python
Copy code
response = client.chat.completions.create(
    model="NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4",
    messages=messages,
    max_tokens=8192,
    temperature=0.7,
    top_p=0.8,
    presence_penalty=1.5,
    extra_body={
        "top_k": 20,
        "chat_template_kwargs": {"enable_thinking": False},
    },
)

Image Input

python
Copy code
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image_url", "image_url": {"url": "https://example.com/image.jpg"}},
            {"type": "text", "text": "Describe this image in detail."}
        ]
    }
]

response = client.chat.completions.create(
    model="NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4",
    messages=messages,
    max_tokens=32768,
    temperature=1.0,
    top_p=0.95,
    extra_body={"top_k": 20},
)

---

⚙️ Recommended Sampling Parameters

---

📥 Download with huggingface-cli

Install the CLI

bash
Copy code
pip install -U "huggingface_hub[cli]"

Download the Full Repository

bash
Copy code
huggingface-cli download NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4 --local-dir ./Qwen3.6-35B-A3B-NVFP4

Download Specific Files

bash
Copy code
huggingface-cli download NeuralNet-Hub/Qwen3.6-35B-A3B-NVFP4 \
  --include "*.safetensors" \
  --local-dir ./Qwen3.6-35B-A3B-NVFP4

---

🔧 Hardware Requirements

[!WARNING] NVFP4 is exclusively supported on NVIDIA Blackwell GPUs. Attempting to run this model on Ampere (A100), Ada Lovelace (RTX 4000), or Hopper (H100) will fail. For those architectures, use the original BF16 model or an AWQ/GPTQ quantized variant.

---

🌐 Contact Us

NeuralNet is a pioneering AI solutions provider that empowers businesses to harness the power of artificial intelligence.

Website: https://neuralnet.solutions Email: info[at]neuralnet.solutions