Wan2.1-VACE-1.3B

Wan2.1-VACE-1.3B is the small build of Alibaba's unified video editing model, published on 13 May 2025 under Apache 2.0 on the same day as its 14B sibling. It does the same three jobs from one set of weights: reference-to-video, where a picture of a subject is kept consistent across a generated clip; video-to-video editing, where an existing clip is rewritten to an instruction; and masked video-to-video editing, where the change is confined to a painted region and the rest of the frame is left alone. The interesting part is what Alibaba had to leave out to make it small. The configuration file shows a 30-layer transformer of width 1536 with 12 attention heads, against 40 layers of width 5120 in the 14B build, and the editing machinery is not spread across every layer: fifteen VACE layers are interleaved at every second position, taking a 96-channel condition input. The weight files count 2.15 billion parameters where the name says 1.3 billion, and the mix is unusual - part of the tensors are stored at full 32-bit precision and part at 16-bit, rather than the uniform format the rest of the family uses. The practical limit is resolution. The 14B build handles both 480p and 720p; this one is recommended at 480p only, because its training at the higher resolution was limited. What it buys in exchange is reach: a model of this size runs on a single consumer graphics card, which is the difference between an editing tool a hobbyist can try and one that needs rented hardware. Usage is modest next to the large build - roughly 4,000 downloads a month for the native weights and 11,000 for the Diffusers conversion published on 4 June 2025, against about 11,000 native downloads for the 14B - so here the smaller model does not out-run the bigger one, unlike in the text and vision lines. Editing appears to be a job people are willing to rent hardware for. The repository includes the LICENSE.txt file its card refers to.

▸Videos

No videos yet.

▸Reviews