# Wan2.1 opened video-generation weights across model size and task variants Alibaba’s Wan team released Wan2.1 inference code and weights on February 25, 2025, with text-to-video and image-to-video models under Apache 2.0. Canonical: https://rangoon.ai/insights/archive-alibaba-wan2-1-open-video-february-2025/ Author: Rangoon Editorial (https://rangoon.ai/insights/editorial/) Published: 2026-10-23 Updated: 2026-10-23 Event date: 2025-02-25 Tags: [Alibaba](https://rangoon.ai/insights/tags/alibaba/), [Wan2.1](https://rangoon.ai/insights/tags/wan2-1/), [video generation](https://rangoon.ai/insights/tags/video-generation/), [Apache 2.0](https://rangoon.ai/insights/tags/apache-2-0/) Source note: Historical reporting: the official Wan2.1 repository records release of inference code and weights on February 25, 2025. It identifies text-to-video 1.3B and 14B models plus 14B image-to-video variants, and states that repository models use Apache 2.0. Later repository additions, such as VACE and first-last-frame work, are outside this February event. Image: Video model variant selection. Original Rangoon explanatory diagram: Source boundary: dated primary release reporting; original editorial analysis. No vendor interface, photograph, measurement result, or endorsement implied. ## Key takeaways - Wan2.1 released multiple task and size variants, so a claim about “Wan2.1” needs the exact model and resolution path. - Apache 2.0 covers the repository models while input-media rights, generated-content use, and hosted infrastructure remain separate decisions. - The 1.3B consumer-GPU option trades capacity and resolution stability differently from the 14B generation paths. ## The February event released code and weights for several video tasks [source 1](https://github.com/Wan-Video/Wan2.1/blob/main/README.md?plain=1) On February 25, 2025, the official Wan2.1 repository recorded release of its inference code and weights. The release included text-to-video models at 1.3B and 14B parameters, with 480P support for the smaller text-to-video model and 480P or 720P paths for the 14B text-to-video model. It also named 14B image-to-video models for 480P and 720P, making task and resolution part of the artifact choice from the start. This is not one universal video generator. Text-to-video starts from a written description, while image-to-video adds a source image whose composition and rights become part of the request. A 480P artifact, 720P artifact, 1.3B checkpoint, and 14B checkpoint can require different memory, time, preprocessing, and quality evaluation, even when a user calls all of them Wan2.1. ## The model family used a video VAE alongside generation models [source 1](https://github.com/Wan-Video/Wan2.1/blob/main/README.md?plain=1) The Wan2.1 repository describes a Wan-VAE that encodes and decodes 1080P video while preserving temporal information, plus separate text-to-video and image-to-video model files. This arrangement makes video generation a pipeline: input preparation, latent representation, model checkpoint, sampling configuration, and video decoding all influence the delivered file. A change in any stage can alter color, motion continuity, resolution, or whether an input image remains recognizable. The repository reported that the T2V-1.3B model needed 8.19 GB of VRAM and could generate a five-second 480P video on an RTX 4090 in about four minutes without optimization. Treat this as a project-reported reference configuration, not a general promise for consumer hardware. Driver versions, offloading, quantization, prompt length, resolution, and concurrent jobs can produce a materially different result. ## Apache 2.0 applies to the released models but not every workflow input [source 1](https://github.com/Wan-Video/Wan2.1/blob/main/README.md?plain=1) [source 2](https://github.com/Wan-Video/Wan2.1) Wan2.1’s repository states that its models are licensed under Apache 2.0 and says it claims no rights over generated content, subject to the license and stated usage restrictions. That is a meaningful artifact-level license statement, and it should travel with the exact code and weights revision selected for a project. It does not grant rights to upload a customer photograph, imitate a person, use a logo, or publish an output that makes an unverified claim. The license also does not define how a third-party host stores prompts or rendered videos. A local download, a cloud notebook, and a managed inference endpoint can create different confidentiality and retention boundaries around the same Apache-licensed weights. Keeping those questions separate lets the release be described accurately without turning “open” into a blanket statement about data handling or creative clearance. ## Task variants make generation controls part of creative review [source 1](https://github.com/Wan-Video/Wan2.1/blob/main/README.md?plain=1) [source 2](https://github.com/Wan-Video/Wan2.1) The official project lists multiple tasks and notes that its 1.3B model has less stable 720P results because of limited training at that resolution. That is a useful example of a capacity tradeoff expressed by the project itself: a model can technically produce an output while still being a poor choice for a required resolution. A creative team should evaluate the desired duration, aspect ratio, motion type, text rendering, and delivery format on the exact checkpoint it intends to use. Later entries in the maintained repository describe Wan2.1 VACE and first-last-frame-to-video additions. They are separate later releases and should not be projected back onto the February 25 code-and-weights event. A historical story gains clarity by naming the original task set, then letting later variants have their own dates, sources, and claims rather than treating a living repository as one frozen feature bundle. ## Sources 1. [Wan-Video GitHub: Wan2.1 release history and README](https://github.com/Wan-Video/Wan2.1/blob/main/README.md?plain=1) 2. [Wan-Video GitHub: Wan2.1 repository and license](https://github.com/Wan-Video/Wan2.1) ## Related reading - [Runway’s Gen-4 focused its March release on keeping visual references coherent across shots](https://rangoon.ai/insights/archive-runway-gen-4-consistent-worlds-march-2025/) - [Google put Veo 2 behind VideoFX’s waitlist before its later cloud release](https://rangoon.ai/insights/archive-google-veo-2-videofx-december-2024/) - [Baidu opened ERNIE 4.5 as a model family, making artifact and license selection part of deployment](https://rangoon.ai/insights/archive-baidu-ernie-4-5-open-model-family-june-2025/) ## More information - [Documentation index](https://rangoon.ai/llms.txt) - [AI access and policies](https://rangoon.ai/ai/) - [Source repository](https://github.com/hypler-dev/rangoon)