Wan model family

Wan is Tongyi Lab’s video stack. The last fully open flagship weights are Wan2.2 (Apache 2.0, July 28, 2025). Everything after that — native audio, 1080P, longer clips — is a closed product line, not a drop-in replacement for the 2.2 checkpoints.

Wan2.2Apache 2.0A14BTI2V-5B
On this page 6
Wan open-source family map from Wan2.1 through Wan2.2 and later product APIs
Open Apache-2.0 weights versus later closed Wan product APIs.
Official Wan2.2 logo
Official Wan2.2 wordmark from the GitHub repository.

What Wan2.2 actually released

The Wan2.2 README lists five downloadable models. T2V-A14B and I2V-A14B are Mixture-of-Experts DiTs (about 27B total parameters, 14B active per step) at 480P and 720P. TI2V-5B is a dense 5B hybrid text/image-to-video model that uses Wan2.2-VAE (4×16×16) and is specified at 720P / 24 fps, including 1280×704. S2V-14B (August 26, 2025) is speech-to-video. Animate-14B (September 19, 2025) is character animation and replacement.

CheckpointTaskSize classResolutionOpen weights
Wan2.2-T2V-A14BText → silent videoMoE 27B / 14B active480P, 720PYes, Apache 2.0
Wan2.2-I2V-A14BImage → silent videoMoE 27B / 14B active480P, 720PYes, Apache 2.0
Wan2.2-TI2V-5BText or image → videoDense 5B + Wan2.2-VAE720P @ 24 fpsYes, Apache 2.0
Wan2.2-S2V-14BImage + audio → talking video14B480P, 720PYes, Apache 2.0
Wan2.2-Animate-14BReference + drive video14Btypically 720PYes, Apache 2.0

What stayed on Wan2.1

First–last-frame (FLF2V) and the all-in-one editor VACE were released on the Wan2.1 line (FLF2V April 17, 2025; VACE May 14, 2025). There is no Wan2.2-FLF2V or Wan2.2-VACE checkpoint in the 2.2 model table. Use Wan2.1-FLF2V-14B-720P and Wan2.1-VACE-1.3B / 14B for those tasks, or later closed APIs that re-expose first/last-frame as a product feature.

Wan 2.5+ is not an open 2.2 upgrade

From Wan 2.5 onward the public story is native audio-video (dialogue, effects, music in one pass), 1080P, and longer duration. Those endpoints are served as APIs. Do not expect Wan2.2-T2V-A14B to emit a soundtrack, and do not treat a “Wan 2.5 open weights” rumor as a substitute for the Hugging Face Wan-AI 2.2 repos.

Official Wan-Bench 2.0 comparison of Wan2.2 against closed models
Official Wan-Bench 2.0 comparison published with Wan2.2.

Why 2.2 exists

  • MoE denoising: a high-noise expert for layout, a low-noise expert for detail, same active compute as a 14B dense step.
  • Cinematic labels: lighting, composition, contrast, color tone in the training mix.
  • More data versus Wan2.1: +65.6% images and +83.2% videos (README).
  • Consumer 720P path: TI2V-5B + Wan2.2-VAE on a 24 GB card with offload.

FAQ

Is Wan2.2 the latest open Wan?
It is the last fully open flagship T2V/I2V generation. Later 2.5–3.0 video models are API products. Separate Apache-2.0 specialists (Animate-2, Dancer, etc.) may still appear; check Hugging Face Wan-AI rather than assuming every version is downloadable.
Does T2V-A14B include audio?
No. Open T2V/I2V output is silent. Audio in the 2.2 family is an input to S2V, not an output of T2V.
Where are the weights?
Hugging Face Wan-AI and ModelScope Wan-AI. Code: github.com/Wan-Video/Wan2.2. Paper: arXiv:2503.20314.

Primary sources

Checked against Wan-Video/Wan2.2 README · Aug 28, 2026

© 2026 wan2.video