Wan

Wan is a family of open-weight video generation models from the Wan Team at Alibaba Group, released under the Apache 2.0 license.¹² The team published the inference code and weights of Wan2.1 on Feb 25, 2025,¹ and its technical report, first posted on Mar 26, 2025, describes a diffusion transformer trained on billions of images and videos, offered in two sizes, 1.3 billion and 14 billion parameters; the 1.3B model needs only 8.19 GB of GPU memory Vendor-stated.² Wan2.1 later added first-and-last-frame-to-video (Apr 17, 2025) and VACE, an all-in-one model for video creation and editing (May 14, 2025).¹ How Wan’s video autoencoder, flow matching and patching work is set out in How AI Video Generation Works: Clip Length, Resolution and Audio by Model.

Wan2.2 followed on Jul 28, 2025. Its A14B models split denoising between two experts, one for the early high-noise steps and one for the later low-noise steps, each of about 14B parameters, for 27B in total with 14B active per step, a mixture-of-experts design; it was trained on 65.6% more images and 83.2% more videos than Wan2.1, and a separate 5B model (TI2V-5B) generates 720P video at 24 fps Vendor-stated.³ Audio-driven Wan2.2-S2V-14B arrived on Aug 26, 2025 and Wan2.2-Animate-14B, for character animation and replacement, on Sep 19, 2025.³ In 2026 the team added Wan-Dancer-14B, which generates dance video from music (Jul 13, 2026), and Wan-Animate-2 (Aug 7, 2026), both Apache 2.0.⁵⁶ The newest generation, wan3.0-video, is offered through Alibaba Cloud Model Studio with clips of up to 30 seconds at up to 1080P Vendor-stated;⁷ no Wan 3.0 weights were listed in the Wan-AI organization on Hugging Face on Oct 18, 2026.⁴ The most-downloaded Wan-AI repository there, Wan2.1-T2V-1.3B in Diffusers format, logged 264,603 downloads in the 30 days to that date Measured.⁴

Wan is also the base of several open real-time talking-head avatar models. Alibaba’s LiveAvatar is a LoRA (a small set of added weights) on Wan2.2-S2V-14B and reports 45 FPS with a 1.21-second time to first frame on five H800 GPUs.⁸⁹ Soul AI Lab’s SoulX-FlashTalk-14B derives its architecture from Wan2.1-I2V-14B and reports 0.87-second start-up and 32 FPS,¹⁰ while its SoulX-LiveAct pairs the 14B Wan2.1 with a 4B audio module and reports 20 FPS on two H100 or H200 GPUs.¹¹ AvatarForcing distills a 1.3B student from a pretrained 1.3B Wan backbone, with a teacher initialized from Wan-S2V, and reports 34 ms per frame;¹² LiveTalk distills OmniAvatar, “a multimodal variant of Wan2.1”, at 20 times less inference cost and latency than full-step baselines.¹³ These speeds are each team’s own figures on different hardware and settings, so they don’t rank against each other Vendor-stated.