Wan2.2: World's First Open-Source MoE Video Generation Model (launch announcement)

TL;DR

Alibaba’s Wan team launched Wan2.2 on 28 July 2025, positioning it as the world’s first open-source MoE-architecture video generation model with explicit cinematic-style control. Three checkpoints ship: a T2V-A14B and I2V-A14B with two-expert MoE (27B total / 14B activated per step, split into a high-noise expert handling early layout and a low-noise expert handling later detail refinement), plus a hybrid TI2V-5B dense model built on a high-compression VAE (4×16×16) capable of 720P@24fps generation on a single consumer GPU. Code, weights, and tech report are released under Apache 2.0, with day-0 ComfyUI and Diffusers integration.

Key claims

First open-source application of the Mixture-of-Experts architecture to a video diffusion model — two experts specialised by diffusion timestep (high-noise / low-noise), 27B total / 14B active, no per-step inference cost increase vs a 14B dense baseline [tweet body].
Hybrid TI2V-5B uses a Wan2.2-VAE with 16×16×4 spatial-temporal compression, supporting both T2V and I2V at 720P / 24fps on a single 4090-class consumer GPU [tweet body].
Training data scaled +65.6% in images and +83.2% in videos versus Wan2.1, with curated aesthetic labels (lighting, composition, contrast, color tone) enabling cinematic-style control [tweet body].
Apache 2.0 release, with code/weights on GitHub, Hugging Face, and ModelScope; ComfyUI and Diffusers integrated on day 0 [tweet body].

Method

The launch tweet is a product announcement; the body of work is the Wan2.2 series itself. The technical novelty is in two places:

Timestep-conditional MoE for diffusion. Standard MoE in LLMs splits experts by token; Wan2.2 instead splits by diffusion step, with one expert specialised on overall scene layout (high-noise regime) and a second on detail refinement (low-noise regime). This makes the routing decision deterministic in the timestep — no learned router, no auxiliary load-balancing loss — and gets the standard MoE benefit (2× parameter capacity per inference-time FLOP) without the standard MoE training pain.
High-compression hybrid VAE for the 5B variant. A 4×16×16 spatial-temporal compression (the temporal axis matches Wan-VAE; the 16× spatial doubles standard VAE compression) shrinks the latent enough that a 5B dense model can produce 720P/24fps video on a 4090.

The accompanying release packs cover the typical “open foundation release” surface: bf16 weights, three model variants (T2V-A14B / I2V-A14B / TI2V-5B), HF Spaces and ComfyUI integrations on day 0, Diffusers PRs, and a multi-modal tech report.

Results

The tweet itself does not report numbers; the accompanying model card claims Wan2.2-T2V-A14B beats leading commercial models across most key Wan-Bench 2.0 evaluation dimensions, and Alibaba’s press release (alibabacloud.com) frames the MoE split as a “~50% reduction in computational consumption” relative to a hypothetical 27B-dense alternative. No filed page yet covers the Wan2.2 tech report itself; this entry is the announcement, not the paper.

Why it’s interesting

Wan2.2 is the first open-source datapoint for timestep-conditional MoE in diffusion — a routing scheme orthogonal to the FFN-experts vs slot-routing taxonomy laid out in MoE Routing Design (where the standard pipeline routes by token via a learned external router). It complements ERNIE 4.5’s ERNIE 4.5 Technical Report modality-isolated routing axis with a diffusion-timestep-isolated axis, and contrasts with Routing-Free Mixture-of-Experts and Path-Constrained Mixture-of-Experts which keep token-routing but tweak the gating mechanism. On the release-packaging side, it’s a fresh datapoint for Open foundation-model releases — the same single-model-series-with-multi-backend-day-0-support convention as Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation and Qwen3-VL-Embedding-8B, but now in the video-foundation slot that had previously been dominated by closed-frontier releases (Veo, Sora). It also positions Wan2.2 as a probable base for the Wan2.7 / Wan-Image / Wan-Animate continued-pretraining lineage now visible in the wiki (Wan2.7-Image — unified model for image generation and editing (Alibaba Wan announcement), Wan-Animate: Unified Character Animation and Replacement with Holistic Replication).