Alibaba's SparkWan checkpoints skip up to 97% of Wan attention
Six checkpoints on Hugging Face take the Wan 2.1 and 2.2 video models to 90%, 95% and 97% attention sparsity, all of them under Apache 2.0.
Key takeaways
- Alibaba has released six SparkWan checkpoints that run the Wan 2.1 and Wan 2.2 video models at 90%, 95% and 97% attention sparsity.
- All six are published under Apache 2.0, according to the Hugging Face model metadata for each repository.
- The three-step 720P build supports 90% to 95% sparsity in three passes, and the two Wan 2.1 480P builds run four passes at 90%.
- The SparkDiffusion paper reports that at 97% sparsity its three-step setup attains slightly better aggregate quality than the strongest 90% baselines on Wan2.1-T2V-14B-720P.
Alibaba's research arm has posted six SparkWan checkpoints that run the Wan video models while discarding between 90% and 97% of attention, and every one is published under Apache 2.0.
The repositories were created on the alibabagroup account on Hugging Face between 07:27 and 08:18 UTC on September 24, 2026, the project README dates the weights release to September 25, and the Hub's model metadata records Apache 2.0 for all six.
What ships is a matched set. SparkWan2.1-T2V-1.3B-480P-0.90Sparsity and SparkWan2.1-T2V-14B-480P-0.90Sparsity run text-to-video at 480P in four steps at 90% sparsity. SparkWan2.1-T2V-14B-720P-0.97Sparsity takes the 14B text-to-video model to 97%, and the three-step 720P build supports 90% to 95% in three passes. SparkWan2.1-I2V-14B-720P-0.97Sparsity covers image-to-video at 97%, and SparkWan2.2-T2V-14B-480P-0.95Sparsity extends the same treatment to the Wan 2.2 line.
The cards state that the task and architecture are retained from the base weights, and name RoLA for learned sparse low-rank attention and CrossDistill for few-step distillation.
SparkDiffusion is the framework around them. Its repository describes it as a video-generation acceleration framework for diffusion transformer models that combines sparse low-rank attention, few-step distillation and custom high-performance operators, adds FP8 quantization with fused kernels, and targets Wan 2.1 and Wan 2.2 across text-to-video and image-to-video. The README claims 200x or more end-to-end inference speedup over the dense multi-step baseline, and the paper title claims up to 265x on a single GPU.
Running it is not a checkpoint swap. The repository requires Linux with a CUDA-capable GPU, Python 3.10 or newer, a CUDA-matched PyTorch, Triton and flash-attn, and inference runs through the repository's own single-case shell wrappers, which load the native Wan VAE and text-encoder files alongside the SparkWan checkpoint.
On quality, the paper's abstract says SparkDiffusion sustains 97% attention sparsity with strong visual quality on long-sequence 720P generation across the Wan 2.1 and 2.2 backbones and both tasks. In its experiments section, the authors report that at 97% sparsity SparkDiffusion attains slightly better aggregate quality than the strongest 90% baselines on Wan2.1-T2V-14B-720P. Those are the authors' own measurements, taken with three-step inference, FP8 quantization and fused kernels, while the released 97% checkpoints are listed as four-step builds. Their latency figures are measured on NVIDIA H100 and RTX 5090 cards.
The nearest comparison already on the shelf is FastVideo's FastH3, which distills MiniMax's 33B H3 model to four steps at 90% sparsity. Hao AI Lab measured a five-second 768p clip at 132.5 seconds on one B200 with dense attention and 16.2 seconds with its sparse four-step build, a factor of 8.16. SparkDiffusion's published numbers come from different GPUs, so the two cannot be placed on one axis yet.
All six checkpoints are on Hugging Face under Apache 2.0, and the framework's code and toolchain are on GitHub.