Wan2.1-T2V-1.3B

Wan2.1-T2V-1.3B was published on the same day as its 14-billion sibling and does a different job. It is not the best video model Alibaba had in February 2025; it is the one that fits. The published figure is 8.19 GB of video memory, which puts it inside almost every consumer graphics card sold at the time, and Alibaba's own measurement was a five-second 480P clip in about four minutes on an RTX 4090, without quantisation or other tricks. The architecture is the same diffusion transformer as the large model, scaled down: 30 layers, 12 attention heads and a hidden width of 1536 against the flagship's 40, 40 and 5120. It shares the Wan-VAE autoencoder and the umT5 text encoder, so prompts in Chinese and English work the same way. The honest limitation is resolution. The model can technically produce 720P, but Alibaba states plainly that it was trained at that resolution only lightly and the results are unstable; 480P is what it is built for. That admission is unusual enough to be worth quoting - most vendors let the higher number stand unqualified. Over eighteen months this became the most downloaded file in the whole Wan family, ahead of every flagship, which is the clearest evidence available that in open video generation the binding constraint is not model quality but the size of the graphics card people actually own.

▸Videos

No videos yet.

▸Reviews