NVIDIA's Sol-H3 Generates Video Faster Than You Can Watch It

NVIDIA released Sol-H3, an optimized inference stack for MiniMax H3 that generates five seconds of video with stereo audio in a median of 1.653 seconds on an 8x B300 Blackwell system. That's roughly three times faster than the clip itself takes to play back.

Built with MiniMax and Reactor, Sol-H3 is open-sourced under Apache 2.0 and available immediately as an API through Reactor.

The Numbers Behind the Claim

Compared to base H3 running its standard 50-step process, Sol-H3's four-step profile cuts generation time dramatically across every clip length tested. A 5-second clip drops from 18.25 seconds to 1.653 seconds, an 11x speedup. A 10-second clip goes from 50.66 seconds to 3.732 seconds, 13.57x faster. A 15-second clip falls from 99.5 seconds to 6.612 seconds, a 15.05x improvement.

The system also scales down. On a single B300 GPU, Sol-H3 still delivers between 9.45x and 14.29x speedups depending on clip length. All figures represent medians of three measured runs, and the comparison uses full generation pipelines rather than isolated component benchmarks, text encoding, denoising, and video/audio decoding are all included in the timing.

How It Actually Gets There

Sol-H3 combines several techniques into one runtime. Dynamic sparse attention requires no retraining and cut attention setup time from 1.206 to 0.285 milliseconds, a 76.4% reduction. Fused kernels handle normalization, positional encoding, and MLP operations together rather than as separate steps. VAE decoding, historically one of the slower stages in video generation, dropped from 7.55 seconds to 0.602 seconds through parallel, batched processing.

The combined effect also frees up roughly 24GB of memory per GPU, real headroom for running additional models or larger batch sizes alongside generation.

Why "Faster Than Playback" Actually Matters

Video generation has functioned as a batch process since it existed: submit a prompt, wait, review the result, adjust, wait again. NVIDIA's own framing treats crossing below real-time playback speed as the more significant milestone here, not the raw speedup number itself.

Once generation reliably outpaces playback, continuous, interactive video generation becomes structurally possible rather than theoretical, prompts can shape footage while a viewer is still watching earlier segments, rather than the two staying separated by a wait.

Competitive Context

This isn't NVIDIA's first pass at optimizing H3, and it isn't the only optimization effort targeting this specific model this month. NVIDIA's own Sol Engine work in early August delivered a 3.95x speedup on 8x GB200 hardware, and a follow-up "Super Acceleration" approach combining a low-resolution draft with lightweight refinement reached over 20x, though with acknowledged quality tradeoffs since it isn't bit-exact with the baseline.

MiniMax released its own "Fast H3 v1" in late August, claiming roughly 14x speedup on Blackwell hardware. fal built H3 Max, a separately post-trained, speed-optimized variant that ranked first on independent quality leaderboards while running dramatically faster than the base model, already covered here. Sol-H3 is the latest entry in this same race, and notably the fastest verified result yet, while explicitly preserving output quality rather than trading it away for speed.

The Signal in the Noise

Multiple well-resourced teams converging on the same optimization target within a single month, NVIDIA, MiniMax itself, and fal, signals real, sustained pressure to solve video generation's speed problem specifically, not just improve output quality. Crossing below real-time playback isn't a marginal win in that race, it's the threshold that actually changes what kind of product gets built on top of the underlying model.

Does faster-than-playback generation change how you think about what AI video tools could eventually be used for, beyond generating individual clips?

The Details

• Release: Sol-H3, built by NVIDIA with MiniMax and Reactor
• Headline result: 5 seconds of 1344x768 video with stereo audio generated in a median 1.653 seconds on 8x B300 Blackwell
• Speedup vs. base H3: up to 15.54x across tested clip lengths and GPU counts
• Techniques: dynamic sparse attention (no retraining required), fused kernels, parallel batched VAE decoding, precomputed AdaLN caching
• Compatibility: any existing MiniMax-H3 few-step LoRA plugs into the same engine
• License: open-sourced under Apache 2.0
• Access: available now via Reactor's API (day-zero release)

Resources & Reads

• NVIDIA's official Sol-H3 technical page
• Try Sol-H3 via Reactor
• BRC: How fal Turned an Open-Weight Model Into the Fastest AI Video Tool Available