Alibaba's Wan team open-sourced Wan 2.1 under Apache-2.0 on Hugging Face and GitHub, shipping a 14B flagship model alongside a consumer-GPU-friendly 1.3B checkpoint (~8.2 GB VRAM) capable of text-to-video, image-to-video, and bilingual visual text rendering in ComfyUI.
Wan
Also known as: Wan 2.1 · Wan 2.2 · Wan 2.5 · Wan 2.6 · Wan 2.7 · Wan 3.0 · Wan2.1 · Wan2.2 · Wan-AI
Plain English
Alibaba's flagship family of video foundation models. While Wan 2.1 and Wan 2.2 became the standard Apache-2.0 open-weight checkpoints for local video generation, Alibaba took Wan 2.5 through Wan 3.0 closed-source and API-only.
Technical Overview
Wan pairs a spatiotemporal 3D causal Wan-VAE (supporting arbitrary-length 1080p encoding via feature caching) with a flow-matching DiT backbone, upgraded in Wan 2.2 to a Mixture-of-Experts (MoE) architecture (27B total, 14B active per step). Starting with Wan 2.5 through Wan 3.0, Alibaba added native single-pass audio-visual synthesis, 30-second multi-shot narrative controls, and omni-reference conditioning while halting open weight releases in favor of proprietary hosted APIs on Alibaba Cloud Model Studio and fal.ai.
Timeline (7)
2025
Release- Paper
The Wan team published Wan: Open and Advanced Large-Scale Video Generative Models, detailing its flow-matching DiT architecture and 3D causal Wan-VAE that chunks temporal compression with feature caches so long 1080p videos decode without exhausting GPU memory.
- Release
Alibaba released Wan 2.2, introducing a Mixture-of-Experts (MoE) video diffusion architecture (
T2V-A14BandI2V-A14B, 27B total parameters with 14B active per step) that routes early high-noise timesteps to a composition expert and late low-noise timesteps to a detail expert, alongside a unifiedTI2V-5Bmodel. - Paper
The Wan team published Wan2.2-S2V-14B and the Wan-S2V technical report, enabling audio-driven cinematic portrait and full-body video generation conditioned on a single reference image and speech or singing audio—later supported natively in stable-diffusion.cpp.
- Release
At the 2025 Apsara Conference, Alibaba unveiled Wan 2.5-Preview with native single-pass audio-visual synchronization (lip-synced dialogue, sound effects, and background music at 1080p)—marking the family's pivot away from Apache-2.0 open Checkpoints toward closed-source, API-only distribution on Alibaba Cloud Model Studio.
2026
ReleaseFollowing the closed-source Wan 2.6 release in December 2025 (15-second multi-shot narrative and reference character consistency), Alibaba launched the Wan 2.7 commercial model suite across Alibaba Cloud Model Studio and partner inference clouds such as fal.ai without releasing downloadable weights.
- Release
Alibaba launched Wan 3.0 as a closed, API-only release on Alibaba Cloud Model Studio, supporting 30-second single-pass 1080p video generation with synchronized audio and omni-reference inputs (images, video, audio, and documents) while leaving Wan 2.2 as the final Apache-2.0 open-weight release in the family.
See also (16)
Open full graphWan
11 more connections
- Mixture of Experts (MoE)Outgoing · Uses
- Video Generation Models (Sora, Veo, Kling, Wan)Outgoing · Part ofIncoming · Mentions
- Chinese AI LabsIncoming · Enables
- CogVideo & CogVideoXIncoming · Alternative to
- stable-diffusion.cppOutgoing · MentionsIncoming · Uses
- CheckpointsOutgoing · Mentions
- fal.aiOutgoing · Mentions
- Hugging FaceOutgoing · Mentions
- Hugging Face DiffusersIncoming · Mentions
- Serverless GPU & Inference CloudsOutgoing · Mentions
- TransformerIncoming · Mentions