News

Faster-Than-Real-Time, Again: OpenVDN's Hybrid Attention Checkpoint for MiniMax H3

12 September 2026 · RenderBob team

OpenVDN's Video DeltaNet checkpoint runs a 14.4-second 768p MiniMax H3 clip in 8 denoising steps, with a community ComfyUI port already live.

Another entry in the race to make MiniMax H3 faster: OpenVDN's Video DeltaNet hybrid-attention checkpoint generates a 14.4-second, 768p clip in 8 denoising steps, and a community ComfyUI port is already available. This is a different mechanism from the Turbo LoRAs covered earlier. DeltaNet-style architectures update a compact internal state incrementally rather than recomputing relationships across the entire sequence at every step. Hybrid attention uses full attention where it matters most and the cheaper mechanism elsewhere. OpenVDN reports faster-than-playback on 8× B200, 11.23 seconds of wall-clock for 14.4 seconds of video. The 8-step checkpoint is what the ComfyUI port exposes. On a single studio card it is a speed win, not a playback-speed guarantee.

The gap between a research checkpoint landing and a usable ComfyUI node has compressed to days. That is the platform's strength, and it is the same open distribution path behind the custom-node security incidents covered earlier this year. An unofficial community port has not been through the same vetting as an official partner node.

For a studio's speed stack, this is another candidate for the fast, iterate-locally tier: draft on the fastest available checkpoint, lock the creative decision, finish on the quality pass. Benchmark candidates like this one against real production shots on a fixed evaluation protocol, confirm the community port comes from a source you are willing to vet, and only then pin it into the standard toolkit.