VideoDeltaNet pushes MiniMax H3 past playback

VideoDeltaNet is a compelling systems-level attack on video generation’s attention bottleneck, though its headline benchmark depends on expensive datacenter hardware and measures denoising rather than complete end-to-end latency.

• –Exact softmax attention remains available for nearby frames, while distant temporal context moves through a recurrent linear-attention branch.
• –The release includes training code, optimized inference code, 50-step and 8-step checkpoints, making the work unusually reproducible for a model-acceleration project. [GitHub repository](https://github.com/OpenVDN/vdn-minimax-h3)
• –The reported 11.23-second result excludes model loading, prompt encoding, VAE decoding, video encoding, and transfer, so production latency will be higher. [Benchmark analysis](https://minimaxh3.cc/guides/openvdn-minimax-h3)
• –Developers should note the 82 GB checkpoint bundle, eight-B200 reference setup, and separate licensing terms for Apache-2.0 code versus MiniMax-derived weights.
• –A community ComfyUI node lowers the integration barrier, but consumer-GPU performance remains distinct from the official benchmark.