TMTPOST — Following its triple AI release last week, Alibaba has open-sourced Tongyi Wanxiang 2.2 (Wan2.2), a film-grade video generation model that sets new benchmarks in efficiency and creative capabilities. The release includes three models: text-to-video (Wan2.2-T2V-A14B), image-to-video (Wan2.2-I2V-A14B), and a unified text-image-to-video model (Wan2.2-TI2V-5B). Notably, the text-to-video and image-to-video models are the industry’s first video generation models based on a Mixture-of-Experts (MoE) architecture. Each of these models features 27 billion total parameters and 14 billion active parameters, combining high-noise and low-noise expert modules. The high-noise modules handle overall video layout, while the low-noise ones refine details—allowing for around 50% savings in computational resources compared to models of similar scale. This architecture addresses the high resource demands typically associated with long-token video generation. Alibaba says Wan2.2 delivers significant improvements in areas such as complex motion synthesis, human interaction, aesthetic quality, and dynamic expression, pushing the frontier of generative video models.