NVIDIA's Sol-H3 makes MiniMax-H3 video+audio faster than playback
Open Apache-2.0 code, three video modes with native audio, and a day-0 hosted API on fal — though a single B300 still needs 13.7s for that 5s clip.
Copy markdown
1.65s for five seconds of video, sound included
On an 8× NVIDIA B300 (Blackwell) system, Sol-H3 turns a prompt into a 5-second, 1344×768, 24fps clip with native stereo audio in 1.653s median — 11.04× faster than 50-step Base H3 Dense, i.e. rendered faster than it plays. Real-time generative video with audio is now something you can put in a loop.
It's open code, not just an endpoint
Sol-H3 is NVIDIA's inference stack for the open-weight MiniMax-H3 model, released Apache-2.0 on GitHub, with text-to-video, first-frame image-to-video, and Reference-to-Video-and-Audio (Ref2VA) modes. You can read and fork the entire serving path — the weights and third-party pieces keep their own licenses.
Three ways in, and a single-GPU tax
No B300 rack? fal hosts MiniMax-H3 pay-per-use as a day-0 partner (5–15s clips up to 2K with stereo audio), and reactor.inc runs a live sandbox. But latency scales with your budget: one B300 takes 13.745s for that same 5s clip and four take 2.918s — 'faster than playback' is an 8-GPU number.
Build this weekend
Ref2VA fuses up to 9 reference images, 3 video clips and 3 audio tracks into one narrated shot — enough for a talking-avatar loop, generative b-roll on demand, or a music-video pipeline where the model keeps pace with the timeline. Prompts and demos ship on the project page.
Elsewhere: the open stack is racing to real-time
Sol-H3 lands a week after MiniMax-H3 reached real-time serving on vLLM-Omni via FastVideo's FastH3 (Sept 1). The open video ecosystem is converging on H3 fast, so expect the single-GPU numbers to keep falling.