Video generation has mostly worked on a batch model so far: submit a prompt, wait somewhere between seconds and minutes, then watch the finished clip. On September 7, 2026, Nvidia Research merged Sol-H3 into the NVlabs/Sana repository — an inference runtime that, on 8× NVIDIA B300 Blackwell GPUs, generates 5.17 seconds of 1344×768 video with native stereo audio in a warm-pipeline median of 1.653 seconds. That is roughly 3.1× faster than the clip plays back — the first time this open-weight stack crossed the real-time threshold on published Nvidia benchmarks.
TL;DR
| Question | Answer |
|---|---|
| What's reported? | Sol-H3 generates MiniMax-H3 video at ~3.1× real-time on 8× B300 |
| What does that mean concretely? | 5.17s of footage renders in 1.653s — faster than playback |
| Why does the threshold matter? | Separates batch generation from live/interactive generation |
| What hardware? | 8× B300 for headline; 4× B300 still ~1.8× real-time; 1× B300 is ~2.7× slower than playback |
| Quality tradeoff? | Four DiT forwards vs 50 in base dense path — speed profile, not attention-only tweak |
| What's excluded from timing? | Model loading, compilation warmup, final MP4 encoding |
| What does this enable? | Live streams, interactive world models, real-time responsive video |
Why "faster than playback" is the meaningful threshold
Most AI video benchmarks report generation time as a multiple of clip length, but the number that actually determines what a product can do is whether that multiple is above or below 1×. Below 1× real-time (generation takes longer than the clip's own runtime), video generation can only ever be a batch process — a user submits a request and waits. Above 1× real-time — which is what Sol-H3 achieves at roughly 3.1× on 8× B300 — a system can, in principle, generate video content continuously, ahead of when a viewer needs it, which is the precondition for live and interactive use cases rather than pre-rendered clips.
This is the same threshold explainx.ai flagged as structurally necessary for interactive world models like Runway's GWM Worlds 2 — a world model that responds to a user's action in real time can't do so if its underlying video generation runs slower than the viewer experiences it. A world model that renders behind real time isn't interactive; it's just a slower version of batch generation with extra steps.
The comparison is not hypothetical. Endless AI TV streams already treat H3-family models as always-on generators — but those pipelines buffer ahead of playback. Sol-H3's published numbers suggest the buffer window can shrink from minutes to seconds on datacenter hardware, which changes what "live" means for product design.
Nvidia's published benchmarks — read the table, not the headline
Nvidia's Sol-H3 project page publishes measured latencies across 1, 4, and 8× B300 GPUs at 5, 10, and 15 second output lengths. All runs use 1344×768, 24 FPS, the same prompt and seed, and reference-free text-to-video-with-audio (T2VA).
| GPU count | 5s output (124 frames) | 10s output (243 frames) | 15s output (362 frames) |
|---|---|---|---|
| 8× B300 — Base H3 dense (50 steps) | 18.250 s | 50.660 s | 99.513 s |
| 8× B300 — Sol-H3 (4 steps) | 1.653 s | 3.732 s | 6.612 s |
| Speedup vs dense | 11.04× | 13.57× | 15.05× |
| 4× B300 — Sol-H3 | 2.918 s | 6.993 s | 12.542 s |
| 1× B300 — Sol-H3 | 13.745 s | 37.813 s | 52.260 s |
Three details matter for builders interpreting this table:
1. Real-time threshold is GPU-count dependent. On 8× B300, every row beats playback. On 4× B300, the 5-second clip still clears at 2.918 s (~1.8× real-time), but longer clips need more headroom. On 1× B300, the same 5-second clip takes 13.745 s — roughly 2.7× slower than playback. Nobody is running consumer-interactive video on one B300 with this profile today.
2. Sol-H3 is not an attention-only patch. Base H3 uses 50 scheduler points (49 DiT forwards); Sol-H3 uses four DiT forwards via the FastH3 Preview v1 adapter. That is a different complete profile — fewer denoising steps — not purely sparse-attention acceleration on the same 50-step path.
3. Timing scope is explicit. Nvidia includes text encoding, DiT denoising, and video/audio VAE decoding. It excludes model loading, compilation warmup, and final MP4 encoding. Cold-start latency for a first user request will be higher than the warm median of three runs after one warmup that Nvidia reports.
How Sol-H3 gets faster — the engineering stack
Nvidia documents six optimization layers on the project page, each with measured evidence. Together they explain why Sol-H3 is more than a distilled checkpoint:
Sol-Attn (sparse attention on the fly). Query-dependent block selection keeps important attention blocks exact and approximates the rest from pooled K/V summaries — up to 2.1× end-to-end versus dense attention pipelines, with no retraining required.
Fused GPU kernels. Residual + RMSNorm + modulation, QKNorm + partial RoPE, and SwiGLU chains stay on-chip instead of round-tripping GPU memory — Nvidia estimates ~178 GB traffic avoided per GPU per DiT evaluation.
8-GPU communication. Ulysses context-parallel exchange with INT8 QKV and FP8 attention output transport cuts communication payload to 51.6% of BF16 on the fastest profile.
Parallel VAE decoding. A 5-second output spans seven temporal clips and 196 spatial tiles; tile sharding plus batched compiled decode dropped per-clip VAE time from 7.55 s → 0.602 s in Nvidia's A/B.
AdaLN precomputation. Conditioning values are projected once before denoising, releasing ~24 GB per GPU and removing ~400 indexing launches per step.
Sol-H3 code ships under Apache 2.0 in the NVlabs/Sana repo; MiniMax-H3 weights and third-party components retain separate license terms — including MiniMax's geographic restrictions on open-weight deployment in the US, EU, UK, and South Korea without authorization, which explainx.ai covered in the MiniMax H3 launch guide.
Where this fits Nvidia's broader inference push
This result sits alongside Nvidia's other 2026 infrastructure announcements aimed at making generative AI viable at consumer-facing latency, rather than research-lab latency. Explainx.ai covered Nvidia's Cosmos 3 open physical-AI world model, which targets the same interactive-simulation category from the model-architecture side; Sol-H3 targets it from the inference-infrastructure side — the chips and serving stack needed to run a video-generating world model fast enough for it to feel responsive rather than laggy.
The timeline converges quickly:
- August 2026: MiniMax announced Fast H3 v1 with a provider-reported ~14× Blackwell speedup; fal Research shipped H3 Max on GB200 NVL72 claiming ~35× vs MiniMax's endpoint.
- September 7, 2026: Nvidia publishes Sol-H3 with reproducible benchmark tables and open runtime code.
- Roadmap: Nvidia states Sol-H3 is moving toward continuous 24 FPS output and direct real-time interaction — prompt updates without restarting the session.
That combination — a world model capable of interactive generation, paired with inference infrastructure fast enough to serve it live — is what turns "AI video generation" from a content-creation tool into infrastructure for new product categories: live-streamed generative content, real-time game or simulation rendering, and responsive world models a user can explore rather than just watch. Unitree's UniFoLM-X2 and robotics simulators face the same latency wall from the physics side — video world models are the graphics half of that problem.
Sol-H3 vs other "faster than real-time" H3 efforts
| Effort | Who | Hardware claim | Method | Public benchmarks? |
|---|---|---|---|---|
| Sol-H3 | Nvidia Research | 8× B300 → 1.653 s for 5s clip | FastH3 4-step + Sol-Engine stack | Yes — full tables on project page |
| Fast H3 v1 | MiniMax | ~14× on Blackwell (provider) | Inference variant | Limited at announcement |
| H3 Max | fal Research | ~35× vs MiniMax endpoint | Post-trained + GB200 serving | fal blog post |
| H3 Super Acceleration | Nvidia SANA team | 6.85 s for 5s 1344×768 on 1× GB200 | 896×512 draft + LTX refinement | Yes — separate project page |
| Sol-Super E2E | Nvidia | 22.2× vs SGLang baseline | Two-stage draft + refine | Published August 2026 |
Sol-H3 and Super Acceleration solve different problems. Super Acceleration uses a low-res draft (896×512, four steps) then upsamples and refines at target resolution — excellent for cost-per-video economics (Nvidia estimated 97%+ GPU gross margin at MiniMax API pricing assumptions). Sol-H3 targets single-pass 1344×768 output with native audio in under two seconds on 8× B300 — closer to what interactive world models need for consistent resolution without a visible draft phase.
What people are asking about Sol-H3
"Can I run this on my RTX 5090?" Nvidia's published benchmarks cover B300 Blackwell datacenter GPUs only. Consumer Blackwell paths are listed as unvalidated in this release. Antirez's H3C port and local MiniMax H3 Max API workflows are separate efforts with different latency profiles — do not extrapolate 3× real-time from datacenter tables to laptop Metal.
"Does four-step generation look as good as fifty?" Nvidia positions Sol-H3 as a speed profile with a dense mode available when behavior closer to the reference model matters. Few-step video distillation always trades something — motion diversity, fine texture, audio sync edge cases. Apply how to read AI benchmarks: compare outputs at your resolution and prompt class, not only wall-clock.
"Is this the same as Seedance or Sora speed claims?" No — Sol-H3 is an inference runtime for MiniMax-H3, an open-weight model with native audio. Seedance 2.5 and closed APIs optimize different stacks. Cross-vendor "fastest video model" headlines rarely share resolution, step count, or hardware — Nvidia's table is unusually explicit on all three.
"What about agentic video pipelines?" When generation drops below playback time, VIMAX-style agentic video loops become practical: generate ten takes, score automatically, pick the best — in the time one take used to consume. Sol-H3 does not include agent orchestration; it removes the bottleneck those agents hit.
"Will Nvidia ship this as a product?" Today it is research code + project page, with an online demo and operator tooling linked from Nvidia's site. Production serving still implies 8-GPU nodes per concurrent stream at headline speeds — a cost model fal's H3 Max economics post already explored for GB200-class hardware.
Builder takeaways — what changes this week
If you ship video products: Treat faster-than-playback generation at your target resolution as a product-requirements milestone, not a benchmark brag. Sol-H3 proves it is achievable on MiniMax-H3 at 1344×768 with datacenter Blackwell — your stack may differ. Profile on your GPU count before promising interactive UX.
If you build world models: Pair model architecture work (Cosmos 3, GWM Worlds 2) with inference runtime work (Sol-Engine, Sol-Attn). The FastGen parallel decoding research Nvidia explored earlier in 2026 is the same thesis: latency is a full-stack problem.
If you run locally: Use Unsloth Desktop or provider APIs for iteration; reserve Sol-H3-class multi-GPU setups for deployment proofs. License geography matters — verify MiniMax weight terms before reproducing Nvidia's benchmark legally.
If you evaluate vendors: Ask for warm vs cold latency, GPU count, step count, and whether audio is included — Nvidia publishes all four; many vendor claims do not.
What's still unverified or narrow
Nvidia's numbers are vendor-published but unusually detailed — not the same as independent third-party replication on different hardware. Gaps to watch:
- No consumer GPU validation in this release
- Warm-pipeline medians understate first-request latency
- Four-step profile vs fifty-step dense — quality comparisons need side-by-side viewing, not only PSNR-style metrics Nvidia did not publish here
- Concurrent streams on one 8× node — headline numbers assume one request saturating the cluster
- Geographic weight restrictions may block legal reproduction in major markets
The video-generation space has had benchmark-methodology disputes before — explainx.ai's guide to reading AI benchmarks applies here: check resolution, clip length, step count, and hardware before treating any speed claim as general capability.
Related on explainx.ai
- AI video generation in 2026: complete guide to Sora, Runway, Kling
- Nvidia Cosmos 3: open physical-AI world model guide
- Runway's GWM Worlds 2: it keeps a world running
- ByteDance's reported Seedance world model, a Genie rival
- VIMAX: agentic video generation, complete guide
- MiniMax Fast H3 v1: Blackwell real-time open video
- fal H3 Max: faster-than-real-time on GB200
- How to read AI benchmarks
Sources
- Nvidia Sol-H3 project page — benchmark tables, methodology, optimization breakdown (September 2026)
- MiniMax H3 Super Acceleration — draft-and-refine alternative path (August 2026)
- NVlabs/Sana repository — Sol-H3 runtime code, Apache 2.0 (merged September 7, 2026)
Benchmark figures reflect Nvidia Research's published Sol-H3 project page as of September 11, 2026. Hardware configs, step counts, and excluded timing phases are documented on that page — verify before citing as reproducible on your stack.