NVIDIA Releases Code That Generates MiniMax Video Faster Than Playback on Eight B300s

NVIDIA has released code that generated a five-second MiniMax H3 video with synchronized stereo audio in less than two seconds—faster than the clip plays. But the result comes from a purpose-built configuration of eight B300 GPUs, four denoising steps, and timing that leaves out loading, compilation warmup, and final MP4 encoding.

The new package, Sol-H3, is NVIDIA’s inference stack for H3, a MiniMax multimodal model that can take text, images, video, and audio as input and generate video with native stereo sound. NVIDIA released Sol-H3’s code under Apache 2.0, so developers can inspect and modify the acceleration layer even though the H3 weights and third-party components remain under separate terms.

A speed result built around fewer model passes

Sol-H3’s largest gain is not simply a faster implementation of the original generation path. It uses a four-step FastH3 adapter that NVIDIA says reduces the workload from 49 transformer forwards to four. The stack also uses approximate sparse attention and INT8 and FP8 data transport.

What the stopwatch includes—and leaves out

NVIDIA timed text encoding, denoising, and video-and-audio VAE decoding. It did not time model loading, compilation warmup, or final MP4 encoding. Those exclusions do not invalidate the measured generation result, but they make it a narrow measure of warm, already-prepared inference rather than an end-to-end delivery time.

Why the headline speed will not transfer automatically

• The same five-second output took 13.745 seconds on one B300 GPU, below playback speed.
• NVIDIA’s fastest profile prioritizes speed with lossy acceleration; a dense mode remains available when behavior closer to the reference model is required.
• The benchmark does not include the work required to load a model or package the finished video file.

Open code does not mean unrestricted weights

The licensing split limits who can reproduce the full stack. MiniMax’s H3 community-license FAQ says open-weight use is temporarily restricted in the United States, European Union, United Kingdom, and South Korea. Organizations in those regions can apply to MiniMax for authorization, while the Apache-licensed Sol-H3 code remains available to inspect and change.

For teams without an eight-GPU machine, Reactor has published a FastH3 API for queuing, generating, and playing H3 clips without users first configuring a multi-GPU machine. It offers a way to use those functions without first setting up the benchmark’s hardware configuration.