RunningHub Open-Sources MiniMax H3 Lightning, Claiming 12x Speedup for Local Video Generation

The reported benchmark uses four RTX 6000D GPUs to generate a 5-second, 1344×768 video. A baseline BF16 50-step run took 348.8 seconds, or nearly six minutes. With the full H3 Lightning acceleration, RunningHub says the same job took 28.7 seconds, reducing generation time by about 92 percent. The output still retains BF16 precision, according to the report.

For longer videos, RunningHub says eight RTX 6000D GPUs can generate a 15-second, 768×1344 video in about 48 seconds from text and about 73 seconds with two reference images. That brings a 15-second image-to-video task to roughly the one-minute scale. QuantumBit's own tests reported about 30 seconds for a 5-second text-to-video clip, about 88 seconds for a 15-second image-to-video clip, and about 50 seconds for a secondary edit of the 15-second clip.

The acceleration stack is described as three engineering layers. First, a RunningHub post-training acceleration model reduces the number of generation steps. The original H3 inference default is 50 steps; RunningHub says a 9-step test scheme cut a 5-second video job on four RTX 6000D GPUs to 43 seconds, about 8 times faster than baseline. Its recommended setting is 4 steps by default, with an 8-step option for difficult tasks such as fast motion or large actions, letting creators choose between speed and stability.

The second layer targets execution efficiency with SageAttention2, Cache-DiT and torch.compile. SageAttention2 speeds up attention computation, Cache-DiT reuses intermediate results across generation steps to reduce repeated work, and torch.compile optimizes execution to lower scheduling and runtime overhead. When combined with the post-training acceleration model, RunningHub says these changes cut the 43-second result further to 28.7 seconds.

The third layer addresses multi-GPU communication. RunningHub tested different parallel strategies for PCIe-based environments without NVLink. It selected TP2 plus Ulysses4 for tensor and sequence parallelism. In an eight-RTX 6000D, PCIe-only configuration, this combination was about 12 percent faster than TP4 plus Ulysses2 and used about 14 GiB less GPU memory, according to the report. The full stack is integrated into the SGLang multimodal_gen inference engine.

RunningHub says H3 Lightning was designed for PCIe multi-GPU environments without NVLink and uses the publicly purchasable RTX 6000D rather than top-tier data-center hardware. The company has published the local deployment process, including environment installation, model download, service startup and inference testing. Developers can also combine the open-source community acceleration LoRA with attention optimization, caching and multi-GPU parallelism on their own machines. For creators without multi-GPU servers, H3 Lightning is also available on RunningHub's platform.

The release follows RunningHub's earlier work around MiniMax H3 after the model was open-sourced. RunningHub integrated H3 into its platform and open-sourced a full set of ComfyUI nodes. QuantumBit reported that more than a thousand creators on RunningHub have open-sourced nearly 10,000 creative workflows around H3, including e-commerce multi-reference video generation, audio-driven digital humans that speak or sing, and comic-drama production workflows that combine storyboards, scripts and audio. H3 Lightning extends that work by pushing acceleration and deployment into the open-source ecosystem rather than only offering an API on top of top-tier compute, according to the report.