A small game studio in a shared office wants a short clip of a character leaping over a crate, rendered in a specific style, without hiring a render farm. Until recently, that meant either a costly cloud subscription or a model locked behind a paid wall.
Alibaba changed that equation in February 2025 by open-sourcing Wan2.1 (万相 2.1), a video foundation model from its cloud unit. The move put a capable text-to-video and image-to-video system into the hands of any developer who can run code.
The interesting part is not just that it is free, but that the license is permissive enough for commercial use, which is where most open releases quietly draw a line.
Key takeaways
- Wan2.1 was open-sourced on Feb 25, 2025 under the Apache 2.0 license, per the project's GitHub repository.
- It ships in 14B and 1.3B sizes; the smaller version runs on a consumer GPU with about 8.2 GB of VRAM (Alibaba).
- The model supports text-to-video and image-to-video, per the official release and GitHub.
- Alibaba reported an 86.22% VBench score at launch, the top mark it cited (company-stated).
- It is downloadable from GitHub, Hugging Face, and ModelScope (魔搭), broadening global access.
Why open-sourcing a video model matters
Closed video models tend to meter usage by the second and cap resolution. Wan2.1 flips that: the weights and inference code are public, so a team can host the model themselves, modify it, and ship derivative products without per-clip fees. That is a structural change, not a discount.
For the broader field, this continues Alibaba's open-weight strategy that began with its Qwen (通义千问) language models. The company has said its Qwen family has spawned more than 100,000 derivative models — a company-stated figure that signals how far open releases can travel once a community forms around them.
Open releases also create a feedback loop. When researchers can inspect the weights, they surface failure modes, propose fixes, and publish them, which raises the whole ecosystem's floor over time. That differs from a black-box API, where the only signal a vendor receives is a support ticket. Alibaba's bet is that a larger community around Wan2.1 ultimately pulls more users toward its paid cloud services, where the real margin sits.
What you actually get
The two sizes serve different users, and the gap between them is the whole story:
- The 14B version targets higher fidelity and supports both text-to-video and image-to-video at 480p and 720p.
- The 1.3B version needs roughly 8.2 GB of VRAM and can produce 480p video on hardware like a consumer RTX card, according to Alibaba's release.
- Both are built on a diffusion transformer architecture with a Flow Matching training approach and a causal 3D VAE for efficient encoding and decoding.
Alibaba reported that Wan2.1 scored 86.22% on the VBench leaderboard at release, which it described as the top result at the time. That number is the company's own claim and was not present in the public GitHub README we reviewed, so treat it as vendor-reported rather than independently confirmed.
The honest trade-offs
Open does not mean easy. Running the 14B model still demands serious compute, and the 1.3B version trades quality for accessibility. The permissive license lowers legal risk for commercial use, but teams must still handle moderation, watermarking, and content safety themselves, which is real work that vendors usually absorb.
There is also a strategic read: by giving the model away, Alibaba competes less on the model itself and more on the cloud compute and enterprise services layered on top. The model is the hook; the infrastructure is the business.
Where Wan2.1 fits in the field
Wan2.1 entered a crowded field of open video models, including SkyReels, CogVideoX, and HunyuanVideo, each with different strengths in clip length, motion control, or licensing terms. Its edge is the combination of a permissive license and a small variant that runs locally, which lowers the barrier for experimentation and teaching.
That breadth is also its limitation. A 1.3B model that fits on a consumer card will not match the motion coherence of the 14B version or of closed systems on hard prompts. Buyers should pick the size for the job, not assume one release covers every use case from social clips to film pre-visualization.
Honest limitations
This article is based on Alibaba's official release, the Wan2.1 GitHub repository, and Chinese securities press reporting. The VBench 86.22% figure is company-stated and was not independently reproduced by us. We did not run the model locally for this piece. Hardware requirements reflect Alibaba's published specifications, which can vary by setup, driver, and quantization choices.
What readers can do now
- Developers should start with the 1.3B Wan2.1 build on a consumer GPU to prototype before renting cloud 14B inference.
- Studios should self-host Wan2.1 to avoid per-second metering on high-volume short-form video.
- Researchers should compare Wan2.1 against closed models on current VBench snapshots, not a single launch claim.