March 4, 2026

How to Deploy Qwen3.5-9B with vLLM

Younes El Hjouji

Detailed Deployment Instructions

1. Environment Setup

# Create a working directory (choose your preferred location)

# Create virtual environment with Python 3.12
python3.12 -m venv venv

# Activate virtual environment
source venv/bin/activate

# Verify Python version (should be 3.12.x)
python --version

# Upgrade pip
pip install --upgrade pip

# Install vLLM nightly (REQUIRED — Qwen3.5 is not yet in stable vLLM)
# Option A: Using pip with nightly index
pip install --pre 'vllm>0.16.0' --extra-index-url https://wheels.vllm.ai/nightly

# Option B: If option A installs stable instead of nightly, force it:
# pip install vllm --extra-index-url https://wheels.vllm.ai/nightly

# Install video support
pip install 'vllm[video]'

2. Launch Deployment

# Make sure you're in your working directory and activate virtual environment
source venv/bin/activate

# Launch vLLM serve
vllm serve Qwen/Qwen3.5-9B \
  --dtype bfloat16 \
  --max-model-len 32768 \
  --max-num-batched-tokens 32768 \
  --gpu-memory-utilization 0.90 \
  --reasoning-parser qwen3 \
  --default-chat-template-kwargs '{"enable_thinking": false}' \
  --port 8000

Note: First model download takes ~5-10 minutes. First inference request triggers CUDA graph compilation which may take several minutes. Subsequent requests are fast.

3. Verification

3.1. Check server is running:

curl http://localhost:8000/v1/models

3.2. Test text inference:

curl http://localhost:8000/v1/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3.5-9B",
    "prompt": "Hello, how are you?",
    "max_tokens": 50
  }'

3.3. Test video inference:

VIDEO_BASE64=$(base64 -w 0 /path/to/your/video.mp4)

time curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d "{
    \"model\": \"Qwen/Qwen3.5-9B\",
    \"messages\": [{
      \"role\": \"user\",
      \"content\": [
        {
          \"type\": \"video_url\",
          \"video_url\": {
            \"url\": \"data:video/mp4;base64,\$VIDEO_BASE64\"
          }
        },
        {
          \"type\": \"text\",
          \"text\": \"Describe the user actions in this video in a sequential list.\"
        }
      ]
    }],
    \"max_tokens\": 1024
  }"

3.4. Test with thinking enabled (per-request):

curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3.5-9B",
    "messages": [{"role": "user", "content": "What is 27 * 13?"}],
    "max_tokens": 512,
    "chat_template_kwargs": {"enable_thinking": true}
  }'

Verified Results:

Text inference: Working. Response instant (warm).
Video inference: Working. video.mp4 (6s clip) processed in ~4.6s.
Video output correctly identified man in red jersey on field, arm gestures, clenching fists, speaking — sequential actions accurately described.
Thinking disabled by default, can be toggled per-request.

4. Troubleshooting

Issue 1: First request takes several minutes

Solution: This is expected. CUDA graph compilation happens on the first request. Subsequent requests are fast.

Issue 2: Requires vLLM nightly

Solution: Qwen3.5 architecture (qwen3_5) is only supported in vLLM nightly builds. Install from https://wheels.vllm.ai/nightly. Note that pip install vllm may install stable 0.16.0 which does NOT include Qwen3.5 support — use --pre 'vllm>0.16.0' to force nightly.

Issue 3: Transformers KeyError: 'qwen3_5'

Solution: vLLM nightly bundles a compatible transformers version. If you see this error, ensure you're using the vLLM nightly venv, not a system Python.

5. Notes

This is the largest dense model in the Qwen3.5 small series — all 9.65B parameters are active per token.
GPU memory for weights (~19.3GB) is moderate; the remaining GPU allocation goes to KV cache for the 32K context.
On H100 80GB, reduce --max-model-len to 16384-24576 and keep --gpu-memory-utilization 0.90.
The model uses a novel hybrid attention mechanism (Gated DeltaNet + Gated Attention) with a 3:1 linear-to-full attention ratio, providing efficient inference on long contexts.
Vision encoder is larger than the 2B/4B variants (27 layers, hidden 1152 vs 24 layers, hidden 1024), matching the 27B model's ViT.
Chunked prefill is automatically enabled with max_num_batched_tokens=32768.