Introduction

Five seconds of world. 1.653 seconds to infer.

That is the sunny little thunderclap from NVIDIA Research’s Sol-H3 page: MiniMax-H3 video with native audio, generated faster than playback on a single 8× NVIDIA B300 Blackwell box. The timeline is already buzzing — see the

, plus notes from
and
.

If H3 felt like a director’s booth, Sol-H3 feels like someone finally greased the booth’s clock.

What Sol-H3 actually is

Sol-H3 is not a brand-new video model. It is a speed-of-light inference stack for MiniMax-H3, built by NVIDIA Research’s Efficient AI Team & Singapore Lab around Sol-Engine × Sol-Attn.

In plain sunshine: same family of H3 video+audio generation, but the runtime unifies the things that usually chew wall-clock time —

  • sparse attention that keeps important blocks exact
  • fused GPU kernels (less memory traffic between ops)
  • faster multi-GPU communication
  • parallel video/audio VAE decoding
  • serving-minded coordination so the whole system stays hot

Showcase clips on the project page lean on Sol-H3’s four-step path (including FastH3’s four-step adapter heritage) at 1344×768 with native audio — not a 50-step crawl pretending to be real-time.

Pro Tip

When you compare “Base H3 Dense” vs “Sol-H3” numbers, remember the profiles differ on purpose: Base uses 50 scheduler points (49 DiT forwards); Sol-H3 uses four DiT forwards. This is a full stack change — not “same steps, sparsify attention, ship it.”

The numbers that made the feed gasp

Same prompt, same seed (20260903), 1344×768, 24 FPS, stereo audio, reference-free T2VA. Median of three runs after one warmup. Timing includes text encoding, DiT denoising, and video/audio VAE decoding — and excludes model loading, compilation warmup, and final MP4 encoding.

On 8× B300, the headline table looks like this:

8× NVIDIA B300 · 1344×768 · 24 FPS · stereo audio
Length | Base H3 Dense (50 steps) | Sol-H3 (4 steps) | Speedup
5s     | 18.250 s                 | 1.653 s          | ~11.04×
10s    | 50.660 s                 | 3.732 s          | ~13.57×
15s    | 99.513 s                 | 6.612 s          | ~15.05×

That 5s row is the vibe: playback is 5s, inference is ~1.65s — past real-time on the measured path.

Attention transport note from NVIDIA: dense attention on 1× B300; on 4× / 8× B300, SOL uses INT8 QKV / FP8 output transport (BF16 transport still available when you want more numerical fidelity).

Why it flies (without the jargon haze)

Think of Sol-Engine as the stage manager and Sol-Attn as the spotlight operator.

Sol-Attn chooses blocks on the fly from the model’s own Q/K/V — query-dependent thresholding, training-free, no weight surgery. Important blocks stay exact; the rest get a cheap approximate correction in the same online-softmax pass.

Around that, Sol-H3 packs the boring-but-deadly overheads: fuse norm/RoPE/MLP-style chains so tensors stay on-chip, pack/all-to-all/merge for eight-GPU exchange without extra full-tensor layout copies, build sparse plans in fewer passes, decode VAE tiles in parallel across GPUs, and precompute AdaLN so denoising does not pay hundreds of tiny launches per clip.

Less waiting. More world. Literally their closing line — and it lands.

Related path, different lane: FastH3 & serving

Don’t mix the nameplates. Sol-H3 is NVIDIA’s end-to-end Sol-Engine stack on Blackwell. Separately, the community serving ecosystem (including FastH3-style adapters and stacks such as vLLM-Omni) is another road toward snappy H3 inference. Same destination energy — faster H3 — different map. Credit where due: NVIDIA’s own acknowledgements call out Hao AI Lab @ UCSD’s FastH3 four-step adapter as the default Sol-H3 path ingredient.

Open, with clear eyes

MiniMax-H3’s open weights still live under MiniMax’s Community License (regional/commercial caveats apply — always read the license for your setup). Sol-H3 itself ships as an NVIDIA Research project-page release with papers for Sol Engine and Sol-Attn; the Sol-H3 page does not loudly brand a single OSI license line the way some repos do, so treat “open” as: inspect the project artifacts and cite the release, don’t assume every file is Apache until you open the LICENSE next to the code you actually run.

Where this is headed

NVIDIA’s research direction is deliciously creator-shaped: continuous 24 FPS chunked generation/streaming with context preserved between chunks, then real-time interactive generation — update a prompt or reference and get the next segment without restarting the session.

That is the glow on the horizon: not just “faster than a coffee break,” but video that keeps up with conversation.

Conclusion

Sol-H3 is the moment MiniMax-H3 stops feeling like a batch job and starts flirting with live. On 8× B300, five seconds of stereo world in about 1.65s of measured inference is the kind of number that rewires what “interactive” can mean for open(-ish) video stacks.

Bookmark the Sol-H3 project page, keep an eye on the

, and when your next H3 render finishes before the preview finishes playing — smile. The booth just got faster.

—Aurelia ♡