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

Wan 2.1 Opens Its Weights, Letting Developers Run Video Generation on a Consumer GPU

Alibaba open-sourced Wan 2.1 under Apache 2.0 in two sizes, including a 1.3B version that runs on a single consumer graphics card.

2026-11-13 · 890 words · NeuroAI
Wan 2.1 Opens Its Weights, Letting Developers Run Video Generation on a Consumer GPU

Open-weight language models have become ordinary. Open-weight video models, until recently, were rare and heavy. Alibaba (阿里巴巴) changed that calculation on February 25, 2025, when it released Wan 2.1 (万相) — the video-generation large model (大模型) inside its Tongyi (通义) family — under the permissive Apache 2.0 license, with the full codebase and weights downloadable by anyone.

The move matters less for the headline "it's open" and more for the size split: one version is large, the other small enough to run at home.

Two sizes, one release

Wan 2.1 shipped in two parameter specifications, and both came with complete inference code and weights:

  • 14B (professional) — the full model for text-to-video and image-to-video at higher fidelity.
  • 1.3B (lite) — a compact version Alibaba says needs only about 8.2 GB of VRAM to produce 480p video, putting it within reach of a single consumer graphics card.

Both tasks — text-to-video and image-to-video — are supported, and the weights landed on GitHub, Hugging Face, and ModelScope (魔搭社区) the same day. For developers used to renting cloud GPUs for every video experiment, the 1.3B variant changes the cost structure of prototyping.

Where Wan came from

Wan is not a sudden entrance. The Tongyi Wanxiang image model first went live on July 7, 2023, and Alibaba unveiled the video-generation version at its Apsara Conference (云栖大会) on September 19, 2024. The February 2025 open release was the point at which the model stopped being a hosted product and became something the community could fork, fine-tune, and run privately.

With Wan open, Alibaba said its open-source line reached what it called "full modality, full size." The claim rests on a long runway: the company says derivative models built on its Qwen (千问) family have passed 100,000 in number, making Qwen one of the largest open model families in the world.

What Apache 2.0 unlocks

The license choice matters as much as the model. Apache 2.0 permits commercial use, modification, and redistribution without a copyleft obligation, which is why Wan 2.1 can be folded into third-party tools, fine-tuned for a specific brand, or run behind a company firewall. For the open community, that lowers the barrier to building wrappers, quantization recipes, and integration with popular node-based interfaces. The practical effect is a feedback loop: more deployments surface more edge cases, and more forks improve the base model faster than a hosted-only product usually allows. Alibaba has leaned on exactly this dynamic with Qwen, where a large derivative ecosystem became a distribution advantage in its own right. Wan 2.1 extends that logic from language to video, a domain where open weights were still scarce at the start of 2025.

What Alibaba says it does well

On quality, Alibaba's own materials focus on a few specifics. The 14B model is described as strong on instruction following, complex motion, physics modeling, and — notably — generating on-screen Chinese text inside video, a capability the team described as a first for the model. In the VBench academic video benchmark, Alibaba said Wan 2.1 reached a total score of 86.22%, which it presented as the top result on that leaderboard at release.

Under the hood, Wan 2.1 uses a self-developed efficient VAE and a DiT (Diffusion Transformer) architecture, with spatiotemporal context modeling that the team says supports 1080p output. Those are architectural claims; the practical takeaway for a developer is that the 1.3B model, while lighter, reportedly came close to some closed models on internal tests while still fitting on commodity hardware.

Why the open release matters

Video generation has been dominated by hosted APIs and a handful of closed frontier systems. Releasing weights under Apache 2.0 inverts that: a studio, a researcher, or a privacy-sensitive enterprise can now run the model on its own servers, modify it, and ship derivatives without per-token fees. For China's open model ecosystem, Wan extends a trend — set by Qwen on language and now by Wan on video — of major labs treating open weights as a distribution strategy, not a risk to be avoided.

Honest limitations

"Open weights" here means the trained parameters and inference code are public; it does not mean the training data, the full training pipeline, or the distillation recipes are shared. The 86.22% VBench score is Alibaba's reported result on one benchmark, and VBench measures specific dimensions (motion, aesthetics, temporal consistency) that do not equal real-world usefulness for a finished film or ad. The 1.3B model is explicitly the lite tier: expect lower fidelity and weaker physics than the 14B version. Like all 2025-era video models, Wan can still produce physically implausible motion, struggle with long clips, and drift from a prompt on fine details. Independent, side-by-side evaluations against Kling, HunyuanVideo, and closed systems are still limited.

What readers can do now

  • Start with the 1.3B model if you have a single GPU. Pull the weights from Hugging Face or ModelScope and test 480p image-to-video before scaling up — it is the cheapest way to learn the model's behavior.
  • Use 14B through a cloud provider for production. If you need 1080p fidelity, rent GPUs or use Alibaba Cloud rather than running the large tier locally.
  • Compare on your own prompts, not the leaderboard. Build a small set of 10–20 clips from your own use case and judge Wan 2.1 against a closed API you already pay for before committing.

Related coverage

Cite this article

NeuroAI (脑机网), "Wan 2.1 Opens Its Weights, Letting Developers Run Video Generation on a Consumer GPU," 13 November 2026, https://neuroai.site/a/na-model-wan-21-open-video

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Foundation Models” → · Back to home · Markdown version

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