Wan 2.2 — implementation by Spacelike AI
MoE video diffusion · 27 B total / 14 B active · Alibaba Tongyi Lab
Wan 2.2, released July 28, 2025, is the first open-source video diffusion model with a Mixture-of-Experts backbone. The T2V-A14B variant routes between a high-noise expert (early denoising / layout) and a low-noise expert (late denoising / detail) based on signal-to-noise ratio, giving it 27 B total parameters while using only 14 B per step.
A companion 5 B dense model (TI2V-5B) pairs a 4×16×16 high-compression 3D VAE with a unified text- and image-to-video head, generating 720p 24 fps clips on a single consumer GPU. Shown here is a single-frame sample from an A14B text-to-video trace.
Specification
- Architecture
- MoE flow-matching DiT · 2 experts · 3D VAE
- Parameters
- 27 B total / 14 B active per step (T2V-A14B)
- Training objective
- Flow matching · SNR-routed expert gating
- Native resolution
- 1280 × 720 · 24 fps · up to 5 s
- Text encoder
- UMT5
- Sampler shown
- Flow-match Euler · 50 steps · cfg 5.0
- License
- Apache 2.0
- Release
- July 28, 2025
- Checkpoint
- Wan2.2-T2V-A14B
Client: Tenstorrent Inc. — performance optimization.
implementation · Hugging Face · Vendor announcement
Live sample on the Spacelike AI home page.
Sample images on this page are licensed under CC BY 4.0 — reuse with attribution to Spacelike AI and a link back to spacelike.ai.