Wan 2.2
Alibaba Tongyi Lab's open-weight video generation family (通义万相 Wan). Apache 2.0 mixture-of-experts checkpoints you can download and fine-tune, separate from the closed Wan 3.0 API.
Overview
Tongyi Wanxiang is the product name. Wan is the code and the weights, and the two have drifted apart badly since mid-2025. Wan 2.2 landed on 28 July 2025 as the first video diffusion model to use a mixture-of-experts split, routing high-noise steps to an expert that handles overall composition and low-noise steps to one that refines detail. Two A14B checkpoints cover text-to-video and image-to-video at 27B total parameters with 14B active per step. A dense TI2V-5B covers both tasks at 720p and 24fps on a single 24GB card, which is the checkpoint most people outside a datacentre will actually run.
Everything in the repository is Apache 2.0. No monthly-active-user threshold, no excluded regions, no clause forbidding you from training other models on the outputs. That licence is why the tooling accumulated here instead of elsewhere: LoRA training scripts, ComfyUI workflows and community fine-tunes for anime, film stocks and specific motion styles. Weights go up on Hugging Face and ModelScope simultaneously, which matters more than it sounds if your build machines sit on a mainland network.
The line keeps moving, but it has split in two. Wan2.2-S2V-14B added audio-driven generation in August 2025, Wan2.2-Animate-14B added character animation in September, and Wan2.2-Animate-2-14B arrived on 7 August 2026 with a distilled 10-step variant quick enough for streaming work. Meanwhile Wan 2.5, Wan 2.6 and Wan 3.0 all shipped as closed commercial APIs, the last on 24 August 2026. Plan your roadmap on the open branch staying roughly one generation behind whatever Alibaba is demoing.
Key Features
- ✓ T2V-A14B and I2V-A14B are mixture-of-experts diffusion transformers: 27B total parameters with 14B active per denoising step, routed between a high-noise expert for composition and a low-noise expert for detail.
- ✓ TI2V-5B is a dense checkpoint aimed at consumer hardware, handling both text-to-video and image-to-video at 720p and 24fps on a 24GB card.
- ✓ Wan2.2-S2V-14B does audio-driven generation, and Wan2.2-Animate-2-14B, published 7 August 2026, does character replacement and motion transfer from a driving video.
- ✓ Animate-2 ships a distillation variant that runs 10 steps with no classifier-free guidance, fast enough for streaming character animation.
- ✓ Every checkpoint is mirrored on ModelScope alongside Hugging Face, with Diffusers-format variants published at the same time.
- ✓ ComfyUI support landed on launch day, and the community LoRA ecosystem for style and motion control is the largest of any open video model.
Where it holds
- • Apache 2.0 with no field-of-use restriction, no monthly-active-user cap and no territorial carve-out, so commercial deployment needs no negotiation with Alibaba.
- • The 5B checkpoint genuinely fits a single 24GB card at 720p, which almost no other video diffusion model manages.
- • Tongyi kept shipping open variants after the flagship went closed: Animate-2 arrived in August 2026, more than a year after the base release.
- • ModelScope mirroring means pulls from inside the mainland do not need a Hugging Face proxy or a VPN.
Where it breaks
- • The open line stops architecturally at 2.2. Wan 2.5, 2.6 and 3.0 are API-only, so native synchronized audio, 1080p and 30-second clips are not in any weights you can download.
- • Official single-GPU inference for the A14B pair asks for 80GB of VRAM. Fitting them on a 24GB card means community GGUF quantization plus offload, and you pay for it in both quality and wall-clock time.
- • Animate-2 defaults are tuned for 8x A800, which is a fair signal of what the team develops against rather than what a solo builder has.
- • English documentation covers inference well but thins out badly for training and fine-tuning, and a good share of the useful issue threads resolve in Chinese only.
My Take
The 5B TI2V checkpoint on a 24GB card is the honest entry point. The 27B MoE pair wants 80GB before you start quantizing. Apache 2.0 with no user cap and no excluded territories is the actual reason this became the backbone of open video work rather than Hunyuan. What you give up is the frontier, because Alibaba shipped Wan 3.0 on 24 August 2026 as a closed API doing 30-second 1080p with synchronized audio while the open branch stayed at 2.2. Still the model I would build a pipeline on, precisely because the licence will not shift under you.
Quick Info
- Pricing:
- open-source
- Openness:
- Open weights
- Licence:
- Apache 2.0
- Starting at:
- Weights are free to download from Hugging Face and ModelScope under Apache 2.0, including commercial use, with no user-count threshold and no separate agreement. Your only cost is compute. The closed Wan 2.5, 2.6 and 3.0 models are billed per generation through Alibaba Cloud Model Studio and are a different product line entirely.
- Added:
- Aug 2026
- Updated:
- Aug 2026
Use Cases
Judge it on your own work
The notes above say where Wan 2.2 holds and where it breaks. The fastest check is your own workload.
Visit website ↗