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

Wan2.1 hit 86.22% on VBench, then Alibaba open-sourced the video model

Alibaba's Wan2.1 topped a major video benchmark and released its weights for free, reshaping what small teams can build.

2026-11-15 · 874 words · NeuroAI
Wan2.1 hit 86.22% on VBench, then Alibaba open-sourced the video model

A graduate student in São Paulo downloads a 14-billion-parameter video model onto a lab server, types a prompt in Portuguese, and renders a five-second clip of a breaking wave without ever opening a paid API console. A year ago that workflow required a corporate contract. Today it requires a GitHub link.

In late February 2025, Alibaba Cloud open-sourced Wan2.1 (通义万相), its visual-generation foundation model for AI video (AI视频), under the permissive Apache 2.0 license. The release put the full inference code and weights of four model variants — T2V-14B, T2V-1.3B, I2V-14B-720P and I2V-14B-480P — on GitHub, Hugging Face and the ModelScope community for anyone to use, modify and ship.

The benchmark that got attention

Wan2.1 arrived already validated by a number. On VBench, the authoritative video-generation evaluation suite, the Wan2.1 series scored 86.22% overall, ranking first among all models tested and standing as the only open-source (开源) entry inside the top five. The score spanned 14 dimensions and 26 sub-dimensions, from motion dynamics and spatial relationships to multi-object interaction.

That placement mattered because it broke a pattern. For most of the previous year, the highest-scoring video models had been closed commercial products behind paywalls. Wan2.1's result told developers that a freely downloadable weights release could still compete with — and in some dimensions beat — the gated alternatives.

Why "consumer GPU" changed the conversation

The detail that lowered the barrier was the 1.3B variant. Alibaba reported that T2V-1.3B needs only 8.19 GB of VRAM, making it runnable on a wide range of consumer graphics cards. On an RTX 4090, the company said, it can produce a five-second 480p clip in roughly four minutes without quantization tricks. For academic labs, indie game studios and solo creators, that is the difference between "we could prototype this" and "we cannot afford to."

The larger 14B models target richer motion and higher fidelity, while the image-to-video variants let a user drop in a single still and a short caption to animate it. A later release, Wan2.1-FLF2V-14B, added first-to-last-frame control: supply a starting and ending image and the model fills the motion between them — a feature aimed squarely at short-form creators who storyboard in stills.

A Chinese-and-English text trick

One capability set Wan2.1 apart from many rivals at launch: it was among the first video models able to render both Chinese and English on-screen text reliably. For a large model (大模型) meant for global use, legible embedded text — a shop sign, a subtitle, a title card — removes a tedious post-production step that previously forced creators back into editing software.

Under the hood, Alibaba built a 3D causal VAE, dubbed Wan-VAE, that encodes and decodes arbitrarily long 1080p video while preserving temporal information, plus a DiT-based flow-matching architecture with a T5 text encoder for multilingual input. These are engineering choices, but they explain why the open release held up under community testing rather than collapsing on harder prompts.

The open-source (开源) math

Alibaba has been on an open-weights streak: it shipped Qwen-7B, its first open large language model, in August 2023, and says it has since open-sourced more than 200 generative AI models, with over 100,000 derivative models built on the Qwen family on Hugging Face alone. Wan extended that posture from text and images into video, the most demanding modality.

The community response was immediate. Within months of the February release, downloads of the Wan2.1 family across Hugging Face and ModelScope crossed 2.2 million, according to Alibaba, and a wave of fine-tunes and wrappers appeared for tasks from advertising to virtual try-on. Open-sourcing did not just hand out a model; it seeded a toolchain.

Beyond the demo reel

The practical upside is not prettier clips but cheaper localization. A small brand can fine-tune the 1.3B variant on its own product photos and generate dozens of language-specific ads without hiring a studio in each market. Chinese cloud providers promptly wrapped Wan2.1 into one-click services, so a team that cannot manage GPUs can still deploy a self-hosted-quality model through a console. Education and virtual try-on were early adopters, where short, templated video is the dominant format and consistency matters more than cinematic surprise.

Honest limitations

This article draws on Alibaba Group's official announcements and Xinhua reporting for the VBench score, VRAM figures, license terms and download counts; I have not independently re-run the benchmark. VBench measures specific technical dimensions and is not a substitute for human judgment of artistic quality or prompt faithfulness. The four-minute RTX 4090 figure reflects an unoptimized baseline and real times vary with resolution and settings. I did not test each variant myself, and the piece does not assess content-labeling (内容标识) requirements for published clips, enterprise support terms, or how Wan2.1 compares with Kling or Vidu on long-form narrative consistency.

What readers can do now

  • If you have a consumer GPU with 8–12 GB of memory, start with the Wan2.1 1.3B variant to prototype video ideas locally before paying for any API.
  • For product or explainer videos needing on-screen text, test Wan2.1's Chinese-and-English text rendering against your current tool to see if it removes an editing step.
  • If you build software, treat the Apache 2.0 license as permission to ship — but confirm the latest model card for any usage restrictions before commercial deployment.

Related coverage

Cite this article

NeuroAI (脑机网), "Wan2.1 hit 86.22% on VBench, then Alibaba open-sourced the video model," 15 November 2026, https://neuroai.site/a/na-aigc-wan-alibaba-open-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