NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

Alibaba Open-Sourced a 720p Video Model That Fits on One Gaming GPU

Alibaba's Wan2.2 put a mixture-of-experts video model in open source, with a 5B variant that renders 720p clips on a consumer GPU.

2026-11-17 · 921 words · NeuroAI
Alibaba Open-Sourced a 720p Video Model That Fits on One Gaming GPU

On a machine built for gaming, a five-second clip appears in under nine minutes. No render farm, no studio pipeline, no waiting in a queue. That is the demonstration Alibaba put in front of developers when it released Wan2.2 (通义万相2.2) in late July 2025.

The interesting part is not the clip. It is what the company chose to hand over: the weights, the code, and a control surface borrowed from cinematography rather than from prompt engineering.

A video model with two specialists

Wan2.2 is Alibaba's second major open-source video family, and its central change is architectural. The larger models — Wan2.2-T2V-A14B for text-to-video and Wan2.2-I2V-A14B for image-to-video — use a mixture-of-experts (混合专家, MoE) design. One network does not handle every step of denoising. Two do: a high-noise expert that fixes the overall layout in the early steps, and a low-noise expert that refines detail later.

The two experts plus shared components hold about 27 billion parameters in total, of which roughly 14 billion activate at any single step. Alibaba says that keeps inference cost and GPU memory close to those of a 14B model while lifting total capacity. The company calls Wan2.2 the first open-source video model built on an MoE architecture. That is a claim from the company's own release, not an independent verdict, and it should be read that way.

  • T2V-A14B: text in, video out, at 480p or 720p.
  • I2V-A14B: a still image in, motion out, at the same two resolutions.
  • TI2V-5B: a compact hybrid that accepts either text or images.

The controls are the point

Resolution is the easy number to quote. The more consequential change is that Wan2.2 tries to let a user specify how a shot looks, not only what happens in it. Alibaba trained the MoE models on a curated aesthetic dataset with labels for lighting, composition, contrast, colour tone, camera angle and focal length. Alibaba's launch materials describe adjustable controls grouped around light, colour and lens language. A user can ask for dusk, a warm palette and a centred composition, and the model is meant to hold those choices across the shot.

That shifts the workflow. A prompt stops being a single sentence you hope the model interprets and becomes closer to a brief you can iterate on. For anyone making storyboards, advertising previsualisation or short-form content, the ability to hold a look steady across takes matters more than a marginal gain on a leaderboard.

Alibaba also reports that, compared with Wan2.1, the new models were trained on 65.6% more images and 83.2% more videos. More data is not the same as better output, but it is a plausible explanation for why the company is pushing motion and semantics harder this time.

The small model is the one that travels

The A14B pair is the headline, but TI2V-5B is the one most likely to end up on ordinary machines. It is a dense 5B model paired with a high-compression VAE that compresses time by 4× and space by 16×16 — far higher than a standard VAE. According to Alibaba's technical blog, it can produce a five-second 720p clip at 24 frames per second in under nine minutes on a single consumer-grade GPU; Chinese technology-press coverage puts the memory requirement at about 22GB of VRAM, within reach of a high-end gaming card.

  • The same model handles both text-to-video and image-to-video, so a creator can move from a written idea to a reference frame without switching tools.
  • Weights and code were posted to GitHub, Hugging Face and ModelScope (魔搭社区) on release day.
  • Alibaba reports cumulative downloads across the Tongyi Wanxiang (通义万相) open-source video series above five million since February 2025. That figure is self-reported and counts the whole family, not Wan2.2 alone.

A benchmark worth waiting on

Alongside the release, Alibaba introduced Wan-Bench 2.0 and reported that Wan2.2 outscored leading closed-source commercial models on it. It is useful as a description of what the team targeted. It is not the same thing as an independent test: a benchmark built by the party being measured establishes a hypothesis. The more reliable signal will be third-party evaluations and, more slowly, what creators actually ship with the model.

There is also a practical limit hiding in the spec sheet. A native generation runs five seconds. A narrative of any length is still a stack of five-second fragments that must be cut, graded and stitched. Better per-shot control reduces the seams; it does not remove them.

Honest limitations

This account draws on Alibaba's official blog and press release and on Chinese technology-press coverage of the launch. I have not run the models, so the speed and memory figures are the company's, measured on its chosen hardware and prompts. The "first open-source MoE video model" wording is Alibaba's own. I have not compared output quality against other commercial video services, examined the licence terms beyond the stated open release, or tested the models on prompts outside English and Chinese. Downstream tools and fine-tunes built on Wan2.2 were outside the scope of this piece.

What readers can do now

  • If you have a GPU with roughly 24GB of memory, download TI2V-5B from Hugging Face or ModelScope and time a five-second 720p clip yourself. Treat the nine-minute figure as a starting point, not a promise.
  • When evaluating any video model, test whether it holds a consistent look across several consecutive shots. That is the part Wan2.2 is designed for, and the part demo reels rarely show.
  • Wait for independent benchmark results before accepting any ranking that comes from a vendor's own test set.

Related coverage

Cite this article

NeuroAI (脑机网), "Alibaba Open-Sourced a 720p Video Model That Fits on One Gaming GPU," 17 November 2026, https://neuroai.site/a/na-aigc-alibaba-wan22-open-source-video

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Generative Media” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories