Alibaba's open-source video generation family, released in late July 2025 under Apache-2.0. Text-to-video and image-to-video, including a 5B variant built to run 720p on a single consumer GPU.
Last verified · 2026-10-26 · by Moe Ameen
Wan 2.2 is the second generation of Alibaba's open-source Wan (Tongyi Wanxiang) video models, released in late July 2025 under an Apache-2.0 license and billed as the industry's first open-source Mixture-of-Experts (MoE) video generation models. Unlike most capable video models in 2026, it is not a hosted-only product: the weights are published on Hugging Face, GitHub, and ModelScope for anyone to download and run.
The family has three models. Wan2.2-T2V-A14B (text-to-video) and Wan2.2-I2V-A14B (image-to-video) are the large MoE models — roughly 27B total parameters with about 14B active per step, splitting denoising between a high-noise expert that sets scene layout and a low-noise expert that refines detail — and Alibaba recommends heavy VRAM (on the order of 80GB) to run them on a single GPU. Wan2.2-TI2V-5B is a smaller dense model that does both text- and image-to-video and was designed to render a 5-second 720p clip in under 9 minutes on a single 24GB consumer card such as an RTX 4090.
What makes Wan 2.2 interesting beyond being open is that it was trained on curated cinematic-aesthetic data, so prompt language about lighting, time of day, color tone, camera angle, composition, frame size, and focal length actually steers the shot rather than being ignored. It integrates with ComfyUI and Hugging Face Diffusers for local use and is also served through Alibaba Cloud Model Studio and third-party hosted endpoints. Note that Wan 2.2 is the open generation that runs on your own hardware; Alibaba's later, hosted Wan3.0 is a separate, longer-clip model reached through Alibaba Cloud rather than a download.
The reason creators reach for Wan 2.2 over a hosted model is ownership: you run the weights yourself, generate at volume with no per-clip meter, and keep the footage private. But that strength creates a specific problem at scale. When you generate forty short clips across a week, each one is an independent 5-second render with no shared identity, no captions, and nowhere to go — a hard drive full of footage, not a brand. [Kompozy](/) is the layer that turns that pile of owned clips into one recognizable, published content operation.
Two things Wan 2.2 deliberately does not do are exactly what Kompozy adds. Identity: a [Persona Brief](/glossary/persona-brief) governs voice and a face-locked AI Influencer persona pool keeps a consistent presenter across every output, so your independently-generated Wan clips stop reading as unrelated one-offs and start reading as the same channel — and [HyperFrames](/glossary/hyperframes) wraps each one in pixel-exact brand styling. Distribution and breadth: Kompozy cuts and reframes each clip into captioned [Clipped Shorts](/glossary/clipped-short) for every feed, and fans the same idea into formats Wan cannot make at all — carousels, quote graphics, a blog, a newsletter, and avatar-led [Persona Shorts](/glossary/persona-shorts) for the weeks you do not render anything. [Autopilot](/glossary/autopilot) then schedules and publishes the whole set across the eight social platforms plus blog and email behind a per-post review gate. You own the model and the footage; Kompozy gives the footage a brand and a destination.
Wan 2.2 is Alibaba's open-source video generation family, released in late July 2025 under an Apache-2.0 license and billed as the industry's first open-source Mixture-of-Experts video models. It has text-to-video and image-to-video A14B models plus a smaller TI2V-5B model that runs 720p on a single consumer GPU, and the weights are downloadable from Hugging Face, GitHub, and ModelScope.
The TI2V-5B variant was built to generate 720p video on a single 24GB consumer GPU such as an RTX 4090, so yes on a capable workstation. The larger A14B text-to-video and image-to-video models need far more VRAM (around 80GB for single-GPU inference), so those are better on a rented high-memory GPU or a hosted endpoint.
The weights are open under Apache-2.0 and free to download, so there is no license fee. Running it still costs your hardware, setup, and electricity, and hosted providers that serve Wan 2.2 bill per generation or per second of output — confirm current rates on the provider, since they vary by resolution.
Wan 2.2 is the open generation you download and run yourself (or reach via third-party hosts), capped at short clips around 5 seconds. Alibaba's later Wan3.0 is a separate, closed, hosted-only model reached through Alibaba Cloud Model Studio and Qwen that generates longer clips (up to about 30 seconds); unlike Wan 2.2, Alibaba has not published its weights.
Bring the clips into Kompozy. It cuts and reframes them into captioned shorts for each feed, and fans the same idea into a carousel, quote graphics, a blog, and a newsletter — holds one face and voice across them, and schedules and publishes the set across the eight social platforms plus blog and email.