Faster than real time
NVIDIA Research has released Sol-H3, an inference stack for MiniMax-H3, the open-weight video generation model. On an 8× NVIDIA B300 Blackwell system, the project page reports that it generates five seconds of 1344×768 video at 24 frames per second, with stereo audio, in 1.653 seconds.
That is faster than the clip takes to play. The comparison the team publishes is against base H3 on the same hardware: 18.25 seconds for the same five-second output, an 11.04× reduction.
The gap widens with length. Ten seconds of video takes 3.732 seconds against 50.66 seconds for the baseline, a 13.57× reduction. Fifteen seconds takes 6.612 seconds against 99.51 seconds, a 15.05× reduction. All of those figures are NVIDIA’s own, measured on its own hardware.
Where the speedup comes from
Sol-H3 combines two pieces of NVIDIA work, Sol-Engine and Sol-Attn, into one runtime. The techniques listed are dynamic sparse attention that requires no retraining, fused norm, RoPE, MLP and sparse-attention setup, fused INT8 QKV and FP8 output communication across eight GPUs, parallel batched VAE decoding, and precomputed AdaLN caching.
The configuration differs by scale: dense attention on a single B300, and the sparse path with INT8 QKV and FP8 output transport on four and eight. The headline numbers also use a four-step generation profile against a 49-step baseline, which is a substantial part of the difference and is stated as such on the project page.
Why an inference stack is the news
The model here has not changed. MiniMax-H3 is the same open-weight system that vLLM wrote about serving in production a week earlier. What changed is everything between the weights and the output.
That is increasingly where video generation economics live. A diffusion video model is expensive per second of output in a way a language model is not, and the difference between a research demo and a product is usually the serving stack rather than the checkpoint. Sol-H3’s code is released under Apache 2.0; model weights and third-party components keep their own terms.
What to watch next
The numbers to look for are independent ones. NVIDIA measured its own optimisations on its own accelerators, which is normal for this kind of release and also a reason to wait for third-party reproductions before treating 15× as a general figure.
The other thing worth watching is whether the techniques transfer. Sol-H3 accepts any MiniMax-H3 few-step LoRA into the same engine, which suggests the engine is not narrowly fitted to one checkpoint — but that claim is easiest to test on a model NVIDIA did not choose.
