The Short Answer by Model
Wan2.2-TI2V-5B is the consumer champion: hybrid text-to-video and image-to-video, 720p@24fps output, roughly 8GB VRAM — it runs comfortably on a laptop-class RTX 4060 or desktop 3060.
The A14B MoE models (T2V and I2V) are the quality flagships. At FP8 scaled precision their two experts load sequentially, peaking under 24GB — an RTX 3090 or 4090 handles them at 720p/81 frames.
FP8 vs GGUF: The Memory Math
FP8 scaled weights run ~14.3GB per expert plus the umt5_xxl text encoder (~6.7GB) and VAE. Sequential expert loading keeps peak VRAM around 22–24GB.
GGUF quantization shrinks each expert to ~10.8GB (Q5_K_M) or ~12GB (Q6_K), bringing total peak into the 12–16GB window via the ComfyUI-GGUF custom node. The trade-off is a modest fidelity reduction — usually acceptable for drafts and social content.
System Requirements Beyond the GPU
Plan for 32GB system RAM when running the 14B experts (16GB works for the 5B model), Python 3.10+ with PyTorch 2.4+CUDA, and 30–50GB of free storage for weights and outputs.
No GPU? Two Free Ways In
First: our online generator runs every Wan version on cloud GPUs — free to try after sign-in, no hardware needed. Second: rented cloud GPUs (~$0.50–3/hour) let you run local workflows without buying hardware. For most creators shipping content regularly, the hosted route wins on total cost once you factor electricity and fiddling time.