MiniMax H3 speeds up with open-community contributions

The FastVideo team, working with Nuva Lab and Nvidia, released FastH3 — a 4-step distillation that runs on DGX Spark and Apple silicon. In parallel, Nvidia's SANA team posted Sol-H3 numbers: 15 seconds of 768p video with audio generated in 6.6 seconds on eight B300 accelerators, measured on their internal warm-inference benchmark.

Haocheng Xi and the OpenVDN group published VDN, a rewritten attention mechanism for faster H3 inference, including weights, training code, and inference code. Alibaba PAI ported Nvidia's PDD distillation method to H3 in an 8-step Acc-LoRAs configuration, now supported in ComfyUI.

LightX2V released 4- and 8-step Turbo LoRAs with workflows covering text-to-video, image-to-video and reference-to-video, all with audio. All components are available in the community repository at github.com/MiniMax-AI/awe.

Researchers and engineers across Nvidia teams, academic labs and independent contributors trained, quantized, tested and shared each release. MiniMax thanked them and invited the community to keep pushing H3 forward.