The Qualcomm AI Research team has introduced MobileWan—an optimized Wan2.2 5B model that allows for high-quality video generation directly on smartphones. Through architectural optimization and distillation methods, the model has achieved record-breaking performance metrics for mobile solutions, setting a new quality standard in the Edge AI segment.

image
image
image

What Happened

Qualcomm AI Research developed MobileWan, an optimized version of the Wan2.2 5B model designed to run on mobile platforms. Key technical improvements include architectural "surgery" involving the removal of approximately 23% of attention heads and the implementation of a hybrid attention mechanism (linear + softmax). The use of Distillation of Diffusion Models (DMD) technology has reduced the denoising process to just 3 steps. As a result, generating a 5-second video (81 frames, 480x832 px resolution, 16 fps) takes approximately 20 seconds. According to the VBench benchmark, the model scored 83.79 points, setting a record for mobile video diffusion models.

Context

Traditionally, heavy video generation models based on the Diffusion Transformer (DiT) architecture require massive computational power from cloud GPUs. The MobileWan project aims to migrate these processes to edge devices, such as smartphones, using aggressive pruning and distillation to reduce computational load without catastrophic loss of quality.

Why It Matters for the Industry

For the industry, this represents a significant technological shift: it proves the feasibility of running complex DiT models on mobile chipsets with acceptable latency. This paves the way for integrating advanced generative functions directly into standard mobile platform SDKs and shifts R&D focus from developing exclusively "large" models toward creating highly efficient architectures for local inference.

Why It Matters for Users

For end users, this means the emergence of applications capable of creating short video clips directly on a phone. Local generation ensures instant response without the latency associated with transmitting data to a server, while also enhancing content privacy, as data never leaves the device.

Sources

Author

Look at AI, Editorial Team