RunningHub open-sources MiniMax H3 acceleration, cutting video generation time about 12x

RunningHub, a one-stop AIGC creation platform, has released and open-sourced H3 Lightning, an inference acceleration stack built around MiniMax's open-source H3 video model. According to the company, generating a 5-second, 1344x768 video on four RTX 6000D GPUs takes about 348.8 seconds under the original BF16 50-step pipeline, but only 28.7 seconds with H3 Lightning - roughly a 12x speedup, or about 92% less generation time. The full stack and local deployment instructions are published on GitHub.

The notable part is what the speedup does not trade away. RunningHub says the pipeline keeps BF16 numerical precision throughout, rather than switching to lower-precision inference or cutting resolution to save time, so quality stays at the same precision tier as the original model.

The work is split into three stages. First, a self-developed RH post-training acceleration model teaches H3 to produce results in fewer sampling steps. The original model defaults to 50 steps, each backed by real computation; with a 9-step test configuration, the same 5-second clip on four GPUs drops from 348.8 seconds to 43 seconds, already about an 8x gain. RunningHub's recommended default is 4 steps, with 8 steps available for difficult cases such as fast motion or large movements, leaving the speed-versus-stability trade-off to the creator.

Second, execution-layer optimizations are stacked on top: SageAttention2 to make attention computation more efficient, Cache-DiT to reuse intermediate results that repeat across sampling steps, and torch.compile to reduce scheduling and execution overhead. Combined with the post-trained model, this pushes the 43-second result down to 28.7 seconds.

Third, the stack targets multi-GPU parallelism. The report notes that on PCIe-only machines without NVLink, inter-GPU communication overhead cannot be ignored. After re-testing several parallelism configurations, RunningHub selected TP2 plus Ulysses4, which on eight RTX 6000D GPUs over PCIe is about 12% faster than TP4 plus Ulysses2 while using roughly 14GiB less VRAM. All of it is integrated into the SGLang multimodal_gen inference engine.

Longer clips scale too. On eight RTX 6000D GPUs, a 15-second 768x1344 video takes about 48 seconds with text-to-video and about 73 seconds with dual reference images. In the author's hands-on test, a 15-second image-to-video run took roughly 88 seconds end to end, and a revision pass took about 50 seconds, bringing 15-second output into the "results in about a minute" range.

Importantly, the optimization targets an accessible hardware profile: PCIe multi-GPU setups without NVLink, using the publicly purchasable RTX 6000D. That is closer to what studios and smaller teams can actually deploy. With a larger budget and cards such as the B300, the report says output time can fall to the seconds range. Creators without multi-GPU servers can also use H3 Lightning directly on the RunningHub platform.

This is also a sample of a wider pattern in open-model ecosystems. After H3 was open-sourced, RunningHub first integrated the model, then open-sourced a full set of ComfyUI nodes; the platform says more than a thousand creators have contributed nearly ten thousand workflows around H3, covering e-commerce video, audio-driven digital humans and animated short dramas. H3 Lightning adds the inference-efficiency layer beneath those application-level workflows.

The signal to watch is that once model weights are strong and freely available, competition tends to shift toward who can turn those weights into stable, reusable production capacity fastest. The report frames this as an extension of parent company Haima Cloud's GPU engineering background - founded in 2014, it has long worked on GPU container scheduling, graphics virtualization and cloud rendering, and now operates more than 60 edge nodes in China serving over 20 million monthly active users. Next to watch: whether the PCIe multi-GPU recipe gets reproduced widely, and how the surrounding acceleration LoRA and workflow ecosystem evolves.