Wan2.2-TI2V-5B

Wan2.2-TI2V-5B is the member of the July 2025 Wan2.2 release that was not designed to win benchmarks. It was designed to fit. Where the flagship of the same generation needs a data-centre card with 80 GB of memory, this one is a five-billion-parameter dense model that Alibaba states will produce a five-second 720p clip in under nine minutes on a single consumer graphics card, naming the RTX 4090, and without special optimisation. The trick is not the parameter count but the compressor sitting in front of it. Wan2.2-VAE squeezes video by a factor of four in time and sixteen in each spatial dimension, a 64-fold reduction, and a further patchification layer brings the total the model actually works with to four by thirty-two by thirty-two. The diffusion model therefore denoises a far smaller representation than its siblings do, which is why a mid-range machine can keep up. Text-to-video and image-to-video are handled by one unified model rather than two, so a single download covers both jobs. What the catalogue can add to the vendor's description is who actually uses these models. Counted over the last thirty days on Hugging Face, this model's diffusers build is downloaded around 190 thousand times, and the single most downloaded Wan repository of all is neither of the 14-billion flagships but Wan2.1-T2V-1.3B, the smallest and oldest of the line, at roughly 242 thousand. The demand in open video generation sits with the models that fit on hardware people already own - which is precisely the segment Alibaba stopped serving when the line went closed after this generation.

▸Videos

No videos yet.

▸Reviews