Wan2.1 put a lab-scale video model under Apache 2.0

The shot

On 25 February 2025, Alibaba's Wan team released Wan2.1, stating in its own GitHub repository that it was releasing "the inference code and weights of Wan2.1," a set of open video generation models rather than a closed API. The repository documents two main sizes, a 14-billion-parameter model and a 1.3-billion-parameter model, released under a public license rather than gated behind a paid product.

What the documents show

The repository states the models are "licensed under the Apache 2.0 License," and a corresponding Hugging Face model card for the 14B text-to-video model confirms the same license and parameter count, "14B params," while noting users are "fully accountable" for their use of the model. The GitHub documentation states the 1.3B model needs "only 8.19 GB VRAM" and can "generate a 5-second 480P video on an RTX 4090 in about 4 minutes," a specific, checkable hardware claim rather than a vague performance statement. The team's own benchmark method, tested against "1,035 internal prompts" across "14 major dimensions and 26 sub-dimensions," states Wan2.1 "consistently outperforms existing open-source models and state-of-the-art commercial solutions" on that internal test; this is the vendor's own benchmark, not an independent evaluation, and no outside party's results are cited to corroborate it.

The workflow

The repository documents the models as usable directly by a developer with local or cloud GPU access: downloading the released weights, running the provided inference code, and optionally fine-tuning the model rather than calling a hosted API. The architecture uses "the Flow Matching framework within the paradigm of mainstream Diffusion Transformers" with a "T5 Encoder" for text understanding, details aimed at a technical audience integrating the model into a custom pipeline rather than an end-user creative tool. The smaller 1.3B model is documented specifically for consumer-grade hardware, while the 14B model targets higher-end setups for improved quality.

What the tool does not change

The repository names a specific limitation: the 1.3B model at 720P is "generally less stable compared to 480P," meaning a production choosing the smaller model still needs to evaluate output quality at the target resolution rather than assume parity. The license places responsibility for lawful use on the operator, not the model, and the documentation does not claim the released weights come with the moderation or content-safety layer a hosted commercial product might include.

• Does the internal benchmark's 14-dimension method test the specific use case a production needs, or only the prompts Alibaba's team selected?
• Does the available hardware match the documented VRAM and generation-time figures for the model size being considered?
• Since the operator bears use responsibility under the Apache 2.0 license, what moderation layer, if any, does a specific deployment need to add?

Wan2.1's release is documented as an access change with real technical detail behind it: openly licensed weights, stated parameter counts, and a self-reported benchmark, rather than an independently confirmed claim of besting every closed competitor.