The Bottleneck Is Named Precisely
What makes this work worth reading is that it starts by saying where the time goes: on MiniMax H3, attention by itself accounts for over 85% of total runtime. Given that, shrinking the model or cutting resolution isn't the efficient move — attention is. Linear attention is of course faster, but on its own it visibly falls down on subject consistency, scene layout and long-range dependencies — the three places video generation gives itself away. VDN-H3's tradeoff runs two branches in parallel: a frame-wise linear one the team calls Video Delta Attention for efficiency, and a softmax branch to hold quality and consistency. The engineering is deliberately restrained — backbone weights are untouched, with only a separate linear attention branch and two small LoRA adapters added and merged at inference time, so it patches onto an unmodified H3 backbone.
Those Eight Cards Only Do Denoising
The 11.23-second figure comes with conditions worth copying down verbatim: eight B200s, 8 denoising steps, 768p, 14.4 seconds of output. The authors also give a deployment note for live setups — run the prompt rewriter, VAE decoding and MP4 conversion on separate machines so the eight GPUs only perform denoising. In other words, this measures the denoising stage alone; pre- and post-processing sit outside it. If you want to try this on one card, the number to compare against is 51 seconds for 8 steps on a single B200. Going from 51 seconds to 11.23 is essentially the eight-way parallel gain, and that curve says more about where this actually sits than the "faster than playback" line does. The release is complete: the 50-step stage-b-step-2000 (4.3GB) and the 8-step stage-dmd-step-250 with turbo adapters (5.1GB), plus MiniMax-H3's 72GB base weights, shipped together with the optimized inference stack and training code. A community ComfyUI port already exists.
Read the License First
This is the part of the story to be careful about: the code repository is Apache-2.0, but the weights come under the MiniMax H3 Community License, and the model card states its scope as worldwide excluding the European Union, the United Kingdom, the Republic of Korea and the United States. That is not boilerplate you can skip — it decides whether a given team can use these weights in production at all. Even outside the excluded list, don't stop at that one sentence; check the original terms for use, redistribution and commercial clauses yourself. License aside, the idea here transfers: rather than reaching for a smaller model, replace the most expensive part of attention and keep a softmax branch to backstop quality. For people working on video generation inference, that structure is worth more study than the timing figure; for people who just want to run it, eight B200s means this is still a cloud proposition.
via: Hugging Face: OpenVDN/vdn-minimax-h3 model card, GitHub: OpenVDN/vdn-minimax-h3, ComfyUI Wiki