Qualcomm research model that runs a distilled, pruned Wan2.2 5B video diffusion model on a mobile device.
Updated EU status checked Last shown
Sources: the vendor's own pages (linked above) and creator videos at the linked timestamps. EU status is information, not legal advice; see how we check.
MobileWan is a Qualcomm AI Research project that deploys a 5B-parameter video diffusion transformer, derived from Alibaba's Wan2.2-5B, on memory-constrained mobile hardware. It converts the model into a chunk-wise autoregressive process with near-constant-memory attention, using recurrence distillation and causal linear attention so inference behaves like an RNN over frames.
Further compression comes from learnable attention-head pruning (23% of heads, using per-head binary gates and a noise-biased sparsity objective), 3-step sampling distillation and a memory-optimized VAE decoder. According to the project page, it generates 5-second 480×832 videos at 16 FPS in about 20 seconds end to end on a commercial mobile device, with a VBench total score of 83.79 (versus 84.16 for the finetuned Wan2.2 5B baseline).
The project page states that models and training recipes will be released; a YouTube creator reported that code and a roughly 9.9GB model have been released, but no repository or model-card links were found in the provided material. The authors list temporal-consistency failures as a known limitation. It is mainly aimed at researchers and developers working on on-device video generation.
Watch the moment: the exact second top creators show it. Videos load only when you press play, via youtube-nocookie.com.
Runs Alibaba's Wan 2.2 5B video model on a phone, producing a 5-second 480p 16fps video in about 20 seconds via chunked generation, pruning, 3-step sampling and an efficient decoder. Code and a 9.9GB model are released.
Closest neighbours by meaning, not just by category.
Model · Alibaba
Alibaba's Wan video generation model, reported to produce clips up to about 30 seconds, available via the Wan platform.
Research · ByteDance
ByteDance and Texas A&M research method distilling image and video diffusion models into few-step generators adversarially.
Model
Research method for multi-shot text-to-video with frame-accurate cuts, fades and camera moves, built on Wan2.1-T2V-14B.
Model
Research model for real-time, talking video agents with a single streaming Transformer; 640×368 output at ~200 ms latency.
Model · Nvidia
NVIDIA's 5B and 14B hybrid linear-attention video diffusion transformers generating up to 720p on a single GPU.
App
Open-source local web app for running video, image, audio, music and 3D generative models on low-VRAM GPUs.