MiniMax and Nvidia researchers this week published Sol-H3, an inference stack that generates AI video faster than the resulting clip takes to play. On an 8x Nvidia B300 Blackwell system, Sol-H3 renders 5 seconds of 1344x768 video with native stereo audio in 1.653 seconds -- down from an 18.25-second unoptimized baseline, and roughly three times faster than the clip's own 5-second runtime. Researcher Enze Xie announced the results directly, and Nvidia's own research page carries the full benchmark: at 10 and 15 seconds of output, the speedup climbs to 13.57x and 15.05x respectively over MiniMax's unoptimized H3 baseline, with single-GPU configurations still landing between 9.45x and 14.29x.
What 'faster than playback' actually measures
- 1.653s · 5-second clip, 8x B300
- Sol-H3 end-to-end generation time
Includes: Text encoding, denoising, and VAE decoding, median of three runs after one warmup run
Excludes: Model loading, compilation warmup, and MP4 encoding -- infrastructure costs the headline number doesn't count - 4 steps · Sol-H3 denoising
- Down from the baseline's 50 diffusion steps
Excludes: Any published visual-fidelity or audio-quality comparison between the two settings
The step count is the detail a headline speed multiple leaves out, and it's a real design choice rather than a pure engineering optimization: fewer denoising passes generally trade some output fidelity for speed in diffusion-based generation, and neither MiniMax's announcement nor Nvidia's technical writeup publishes a side-by-side quality comparison between the 4-step fast path and the 50-step baseline. Nvidia's page is at least explicit that the numbers are a runtime measurement, not a quality claim -- the Sol-Engine code itself is released under an Apache 2.0 license, while the underlying MiniMax-H3 model weights carry MiniMax's own, more restrictive community license.
Sol-H3 is an optimization layer over MiniMax's underlying H3 model, which the company released July 31 as a general-purpose, omni-modal generator: up to 15 seconds of video at 2K resolution with native 32kHz stereo audio, built to follow multi-step instructions across text, image, video and audio inputs at once. MiniMax's own pricing claim -- less than a third of "mainstream models'" per-second cost at 2K, less than half at 768p -- was never independently verified at the time of that release and remains a company figure rather than a measured one; Sol-H3 is the first outside benchmark attached to the H3 line since.
- MiniMax's new Sol-H3 inference stack generates a 5-second AI video with audio in 1.653 seconds on an 8x Nvidia B300 system.
- That's faster than the clip's own runtime, and up to 15.54x faster than MiniMax's baseline H3 inference at longer lengths.
- The underlying H3 model, released July 31, generates up to 15 seconds of 2K video with native stereo audio.
- Caveat: the speed benchmark uses 4 denoising steps versus the baseline's 50 -- a real quality tradeoff neither party has published fidelity numbers for.