Sol-Engine × Sol-Attn · MiniMax-H3 Inference

Sol-H3 Speed-of-Light MiniMax-H3 on an 8× NVIDIA B300 Blackwell System

NVIDIA Research, Efficient AI Team & Singapore Lab.
5sof world. 1.653sto infer.

Sol-H3 generates MiniMax-H3 video with native audio faster than playback on one 8× NVIDIA B300 system. Sol-Engine × Sol-Attn unifies sparse attention, fused kernels, fast multi-GPU communication, parallel video decoding, and other serving optimizations in one runtime.

Sol-H3 code Sol-Engine page Real-time demo
Full-frame preview 1344×768 / 24 FPS
Featured sample · MiniMax-H3 Base Ref2VA at 50 steps, generated through Sol-Engine on 8× NVIDIA B300.
01
Benchmark · 8 / 4 / 1× NVIDIA B300

Latency across GPU counts.

Choose an output length, then compare Sol-H3 directly with Base H3 across 8, 4, and 1 B300 GPUs.

Measured generation time · seconds

Generation time across GPU counts.

Measured latency for 50-step Base H3 Dense and four-step Sol-H3 at 5, 10, and 15 second output lengths on one, four, and eight NVIDIA B300 GPUs.
GPU count 5s124 frames · 1344×768 · 24 FPS · stereo audio
8×NVIDIA B300
Base H3Dense · 50 steps18.250 s
Sol-H3Ours · 4 steps1.653 s11.04× faster
4×NVIDIA B300
Base H3Dense · 50 steps35.328 s
Sol-H3Ours · 4 steps2.918 s12.11× faster
1×NVIDIA B300
Base H3Dense · 50 steps129.898 s
Sol-H3Ours · 4 steps13.745 s9.45× faster
Measured scope All runs use 1344×768, 24 FPS, the same prompt and seed 20260903, and reference-free T2VA with stereo audio. The complete profiles differ: Base H3 uses 50 scheduler points (49 DiT forwards), while Sol-H3 uses four DiT forwards—not an attention-only runtime change. Sol-H3 uses Dense attention on 1× B300 and SOL with INT8 QKV / FP8 output transport on 4× / 8× B300. Each value is the median of three measured runs after one warmup. Timing includes text encoding, DiT denoising, and video/audio VAE decoding; model loading, compilation warmup, and final MP4 encoding are excluded.
02
Selected Sol-H3 generations

Sol-H3 video showcase.

Browse ten-second Sol-H3 videos with native audio, two examples at a time.

Loading showcase Generated with Sol-H3

Selected examples.

Browse two videos at a time with the arrows; each video includes its exact prompt.

Browse examples
Preparing examples Page — / —
Example — Preparing render
Sol-H3 · 4-step
Category — Language — Shots — Seed —
Preparing video —
View exact prompt Official MiniMax-H3 T2VA schema 
Loading…
Example — Preparing render
Sol-H3 · 4-step
Category — Language — Shots — Seed —
Preparing video —
View exact prompt Official MiniMax-H3 T2VA schema 
Loading…
Our prompt manifest ↗ Official H3 prompt skill ↗
Showcase profile Each ten-second sample is generated by Sol-H3 at 1344×768 with native audio, using Sol-H3’s four-step Dense/Data-Free adapter from FastH3 Preview v1.
03
Two core systems

Sol-Engine × Sol-Attn.

Sol-Engine speeds up the complete inference pipeline. Sol-Attn selects attention blocks while the model runs, with no retraining.

01 · Sol-Engine

Make the whole system fast.

Sol-Engine keeps the model ready and coordinates GPU kernels, communication across eight GPUs, compilation, and video/audio decoding.

02 · Sol-Attn

Select the blocks that matter.

Sol-Attn uses a query-dependent threshold to keep important attention blocks exact and cheaply approximate the rest. It works directly from the model’s Q/K/V tensors, so existing weights need no retraining.

04
Six optimizations · one runtime

How Sol-H3 gets faster.

Each card pairs one optimization with how it works, why it is faster, and the measured evidence.

GPU kernels

Fuse repeated operations

Norm · RoPE · MLP
02GPU kernels

Move less data between operations.

Mechanism

Sol-H3 fuses three repeated chains: residual + RMSNorm + modulation, QKNorm + partial RoPE, and SwiGLU. Their intermediate tensors stay in registers or on chip instead of returning to GPU memory between operations.

Why it is faster

These operations touch wide tensors but do relatively little math, so memory traffic is often the limit. Fewer reads, writes, and launches save traffic across all 50 transformer blocks.

Multi-GPU communication

Faster 8-GPU data exchange

Pack · all-to-all · merge
038-GPU exchange

Pack, exchange, and merge directly.

Mechanism

Ulysses redistributes sequence shards by attention head, then restores the original layout. Sol-H3 packs Q/K/V directly for the destination GPU, performs one all-to-all exchange, and merges returned heads without extra full-tensor layout copies.

Why it is faster

This removes separate stack, permute, and contiguous copies. The fastest profile sends QKV in INT8 and attention output in FP8; BF16 transport remains available when higher numerical fidelity is preferred.

Sparse-attention setup

Fuse the setup work

K/V summaries · threshold · metadata
04Sparse setup

Build the sparse plan in fewer passes.

Mechanism

One preprocessing kernel computes each key centroid and value sum together. Thresholding also returns the pooled query for reuse; the final compaction step applies threshold, local-band, and exact-prefix rules while emitting the mask and compact block lists.

Why it is faster

The default SOL path feeds its fused K/V summaries, mask, and block list directly into cuDNN block-sparse attention. The optional native Sol-Attn kernel combines routing and sparse attention in one pass. Both reduce temporary traffic and launch overhead.

Video decoding

Parallel VAE decoding

Tile parallel · batch · compile
05VAE decoding

Decode spatial tiles in parallel.

Mechanism

A 5s output has seven temporal clips and 196 equal spatial tiles. The default path distributes the full tile set over eight GPUs, batches each GPU’s share into one compiled decoder call, and gathers once.

Why it is faster

Tile sharding cuts the work per GPU, while global batching replaces seven per-clip decode-and-gather rounds with one. A per-clip path remains available when matching the initial output exactly matters.

Conditioning

Precompute AdaLN

Precompute · cache · direct lookup
06AdaLN cache

Compute conditioning once, then reuse it.

Mechanism

All AdaLN values are knowable before denoising. Sol-H3 projects every block-and-step value up front, caches the full trajectory, and releases the projection weights before the denoising loop.

Why it is faster

Each block reads its cached slice directly instead of running two index selections. Across 50 blocks and four steps, that removes about 400 small launches and avoids roughly 26 GB of GPU-memory reads per step without changing output.

Measured T2VA 5 / 10 / 15 seconds 1344×768 · 24 FPS INT8 QKV · FP8 attention output 8× NVIDIA B300
05
Future Plan

Where Sol-H3 goes next.

Sol-H3 is moving toward continuous 24 FPS output and direct real-time interaction.

Research direction Continuous generation

24 FPS live streaming

Generate and deliver video with native audio chunk by chunk, preserve context between chunks, and target smooth 24 FPS playback.

Generate → decode → stream
Coming next Interactive generation

Real-time interactive

Update a prompt or reference input and see the next generated segment respond without restarting the session.

New input → next video segment
Credits

Acknowledgements

We thank the teams and open projects that helped turn Sol-H3 from individual optimizations into one end-to-end release.

Core contributors Jincheng Yu Junsong Chen Yitong Li Haopeng Li Haocheng Xi Song Han Enze Xie
  • MiniMaxMiniMax-H3 open weights and code.
  • ReactorHosted online demo and deployment collaboration.
  • Humanize + KDAOperator and fused-kernel optimization on top of Sol-Engine.
  • LightX2VOpen MiniMax-H3 Turbo LoRAs and the Ref2VA workflow used by Sol-Engine.
  • Video DeltaNet · UC BerkeleyComplementary hybrid-attention work for MiniMax-H3.
  • Hao AI Lab @ UCSDFastH3 four-step adapter used by default in Sol-H3.
How to cite

Citations

If Sol-H3 helps your work, please cite this release together with the underlying Sol-Engine and Sol-Attn papers.

BibTeX · Release + core papers
@misc{solh3_2026,
  title        = {Sol-H3: Speed-of-Light MiniMax-H3 on an 8× NVIDIA B300 Blackwell System},
  author       = {{NVIDIA Research, Efficient AI Team \& Singapore Lab}},
  year         = {2026},
  howpublished = {\url{https://nvlabs.github.io/Sana/Sol-Engine/Sol-H3/}},
  note         = {Project page}
}

@misc{li2026solvideoinferenceengine,
  title         = {Sol Video Inference Engine: Agent-Native Full-Stack Acceleration Framework for Efficient Video Generation},
  author        = {Yitong Li and Junsong Chen and Haopeng Li and Haozhe Liu and Jincheng Yu and Ligeng Zhu and Ping Luo and Song Han and Enze Xie},
  year          = {2026},
  eprint        = {2606.23743},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2606.23743},
  url           = {https://arxiv.org/abs/2606.23743}
}

@misc{li2026solattn,
  title         = {Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification},
  author        = {Haopeng Li and Yitong Li and Junsong Chen and Tian Ye and Haozhe Liu and Jincheng Yu and Duomin Wang and Ruihua Zhang and Zeke Xie and Enze Xie and Song Han},
  year          = {2026},
  eprint        = {2607.24027},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  doi           = {10.48550/arXiv.2607.24027},
  url           = {https://arxiv.org/abs/2607.24027}
}
Sol-H3 · MiniMax-H3 Inference

Less waiting.More world.