HunyuanImage 3.0 Hardware Requirements

Model variants and published VRAM

The family uses an 80B-total mixture-of-experts architecture with 13B parameters active per token. “13B active” does not mean only 13B parameters must be stored. The official table should remain the planning baseline unless you have measured a specific community quantization or offloading setup.

Software environment

Tencent's repository currently recommends Python 3.12 and CUDA 12.8, with matching PyTorch packages. FlashInfer and FlashAttention are optional performance components and add compatibility requirements of their own.

pip install torch==2.8.0 torchvision==0.23.0 torchaudio==2.8.0 \
  --index-url https://download.pytorch.org/whl/cu128
pip install -r requirements.txt

Do not copy those pinned versions blindly into an existing environment. First confirm the host's NVIDIA driver and CUDA compatibility, then follow the current commands in the official repository.

Official Instruct setup outline

1. Clone the official repository.
2. Install the documented dependencies in a dedicated environment.
3. Download the checkpoint to a local directory whose name does not contain dots.
4. Point MODEL_PATH to that directory and run the provided script.

git clone https://github.com/Tencent-Hunyuan/HunyuanImage-3.0.git
cd HunyuanImage-3.0

hf download tencent/HunyuanImage-3.0-Instruct \
  --local-dir ./HunyuanImage-3-Instruct

export MODEL_PATH="./HunyuanImage-3-Instruct"
bash run_demo_instruct.sh

What the distilled checkpoint changes

The Instruct-Distil checkpoint is designed for fewer sampling steps; Tencent recommends 8. That can reduce inference time. The official model table still lists ≥8 × 80GB VRAM, so “distilled” should not be translated into “fits on one gaming GPU” without measured evidence.

Cloud deployment checklist

• Choose a provider that can supply the required number of 80GB GPUs in one compatible instance or cluster.
• Verify interconnect, disk capacity, container image, CUDA version, and total hourly cost.
• Use provider pricing shown at purchase time; do not rely on a fixed price copied from an old article.
• Set a spending limit and stop or terminate the instance after the test.
• Start with the official demo before adding quantization, offloading, or a web UI.

If that setup is disproportionate to your use case, compare smaller routes in the model quickstart or use the hosted Nano Banana 2 API. They solve different infrastructure problems.

Primary sources

• Tencent HunyuanImage-3.0 repository: environment, commands, and VRAM table
• HunyuanImage-3.0-Instruct model page