Wan 2.2
The fully open flagship of video generation
What it is
Alibaba's open video suite built on a mixture-of-experts diffusion architecture: 27B text-to-video and image-to-video models, a 5B dense variant with a high-compression VAE for consumer GPUs, plus speech-to-video and character-animation variants.
Why it's interesting
The strongest fully Apache-licensed video family in mid-2026 — photorealistic humans, an unusually broad task spread, and a 5B variant that genuinely runs 720p24 on a single RTX 4090. Its successor (Wan 2.5+) went API-only, making this the open high-water mark.
Use cases
- Text- and image-to-video up to 1080p
- Character animation and replacement
- LoRA fine-tuning for style
Who it's for
Video artists, researchers, studios needing commercially safe open weights
Setup
Advanced. The 5B variant wants a 24GB GPU; the 14B-class variants want 80GB+ or multi-GPU; quantized community builds exist
Limitations & cautions
Clips are seconds-scale, the big variants effectively need datacenter hardware, and the 5B takes minutes per clip unoptimized.
Editorial takeaway
When the next version goes closed, the last open one becomes infrastructure. Expect this checkpoint to be built on for years.