Wan 2.1

Wan 2.1 is Alibaba's open-source video generation model, released February 2025 under the Apache 2.0 license. Built on a diffusion transformer with a 3D causal VAE (Wan-VAE) for better scene consistency, it topped the VBench leaderboard at launch with 86.22%, ahead of OpenAI's Sora and Luma at the time. It ships in 1.3B and 14B variants — the 1.3B model needs only about 8.19GB of VRAM, making it runnable on consumer cards like the RTX 3060. It supports text-to-video and image-to-video at up to 1080p, with English and Chinese prompting, and is widely deployed via Diffusers, ComfyUI, and Docker. Wan 2.2 (July 2025) added the first open-source MoE video architecture; note that 2.5+ versions are API-only commercial models.

The honest picture: Wan 2.1 is a 2025-era model, and its quality trails current closed leaders on fine detail, complex motion, and prompt adherence. The 14B variant wants 24GB+ VRAM, so most users run the smaller model or rent cloud GPUs. Community reports note occasional NSFW-capable forks ('Spicy' variants) that you should steer clear of in professional contexts. Later Wan versions went closed-source, so the open line effectively ends at 2.2.

pip install diffusers and load the WanPipeline from Hugging Face (Wan-AI org), or use ComfyUI Wan nodes / Docker images. 1.3B checkpoint for consumer GPUs.