Day One Optimization

Deploy MiniMax-H3 with Sol Engineachieving 3.95× acceleration in 4.5 hours of optimization

The Sol Video Inference Engine delivers a 3.95× end-to-end speedup on 8× NVIDIA GB200 at near-lossless quality. It gets there without distillation, without LoRA or fine-tuning, and without an offline calibration pass.

01 — MiniMax-H3

One transformer that hears, sees, reads — and answers in video with sound.

MiniMax-H3 (Hailuo 3.0) is a 33B dense omni-modal generation model that takes text, images, video, and audio as a single unified context and returns a 4–15 second clip at up to 2K/24 fps with native 32 kHz stereo audio — dialogue, effects, and room tone synthesized in the same forward pass as the pixels, not dubbed on afterwards. That fusion is what makes it interesting and what makes it expensive: attention runs over a joint video-plus-audio sequence long enough to dominate the profile, roughly 13B of the 33B sits in AdaLN branches that fire on every denoising step, and the reference implementation ships in BF16 written for clarity rather than for speed.

T2VA · text → video10.1 s
I2VA · image → video8.0 s
Ref2VA · reference → video5.2 s

02 — Sol Engine

3.95× end-to-end optimization

Measured on 8× NVIDIA GB200 · 1344×768 · 24 fps · 124 frames.

The Sol Video Inference Engine converged on a recipe that generates the same clip 3.95× faster with no perceptible change in picture or sound. It gets there without any of the machinery that usually buys speed at this scale: no distilled student, no LoRA, no fine-tune, and no offline calibration pass over a dataset. Everything below is inference-time only and runs against the stock checkpoint.

Where the speedup comes from

Each technique multiplies onto the one before it, and together they compose to the final end-to-end number.

End-to-end wall time as each technique is added
Diffusers baselinereference implementation, 8 GPUs
27.205s1.00×
+ Kernel losslessfusion and graph capture, incl. VAE
19.514s1.394×
+ Sparse attentionSol-Attn
17.744s1.534×
+ Cross-step cacheresidual reuse
6.883s3.95×
SGLang baselinesecond reference point
19.300s1.410×
0wall time, warm27.2 s
Bar length is measured wall time on the hot path; the label at each tip is the cumulative speedup against the Diffusers baseline. Below the dashed rule sits SGLang as a second reference point — our final configuration is 2.80× faster than it.
Reference · Quality preserving · Super efficient LPIPS 0 · 0.210 · 0.293

The recipe ships in two configurations. Quality preserving runs the cross-step cache with sparse attention off; super efficient turns Sol-Attn on as well, buying a further 1.25× at a measurably larger perceptual distance. Every clip is its own file with its own audio track — play across a row to compare sound as well as picture. Measured at 1344×768, 124 frames, 50 steps on 8× NVIDIA GB200.

Prompt 1

Cinematic push-in on a starship bridge: a captain in a high-collared navy tunic stands silhouetted at the observation window as a dreadnought armada charges its hyperdrives. A blinding flash, the bridge shudders, and the fleet is gone. [full prompt compacted]

Reference1.00×
Quality preserving3.16×
Super efficient3.95×
Prompt 2

A flamenco dancer spins fast on a wooden stage under a single warm spotlight, her red skirt flaring wide, heels striking the boards in sharp rapid bursts, a guitar driving underneath and hands clapping on the offbeat, camera circling with her.

Reference1.00×
Quality preserving3.16×
Super efficient3.95×
Prompt 3

Two mechanics in a garage at night lean over an open engine bay. The older one wipes his hands and says “It’s not the pump.” The younger one answers “Then what is it?” A radio murmurs in the corner, a wrench clatters onto the concrete, fluorescent tubes buzz overhead.

Reference1.00×
Quality preserving3.16×
Super efficient3.95×

Pillar 1 — Lossless kernel optimization

Before approximating anything, take the free money. A reference implementation written against PyTorch eager spends a startling share of its wall clock on HBM round-trips between small elementwise kernels and on launch overhead.

Fusing those operations is what removes it: each merged kernel is one launch instead of several and keeps its intermediates in registers and shared memory rather than writing them out to HBM and reading them straight back.

FusionWhat it eliminates
fused RMS-AdaLNHBM round-trips across the modulation branch — the largest single win on H3, given its 13B AdaLN share
fused QKNorm + RoPETwo full read/write passes over Q and K, plus 3D MM-RoPE indexing overhead
QKV mergeThree skinny GEMMs collapsed into one wide, tensor-core-friendly GEMM
GEMM + GELUAn activation-sized HBM round-trip per MLP block
torch.compile captureResidual launch overhead and leftover elementwise chains

Pillar 2 — Cross-step cache

Across a denoising trajectory, consecutive steps produce transformer outputs that are extremely similar through the middle of the schedule: the model is refining, not deciding. Cross-step caching reuses a previous step's residual instead of running the full denoiser, gated by a cheap change detector so that the steps which do matter — early structure, late detail — are never skipped.

We use First Block Cache for that gate. It watches the residual produced by the first transformer block and treats it as a proxy for the whole stack: if that residual has barely moved since the last step, the remaining blocks are skipped and the previous step's output is reused. One scalar — the residual-difference threshold — sets how similar counts as similar enough, which is the knob the two presets differ on.

Pillar 3 — Sol-Attn, training-free sparse attention

At high resolution and long duration, with audio tokens interleaved into the same attention, sequence length dominates the cost. Attention maps in video diffusion are also extremely sparse: most query blocks attend meaningfully to a small fraction of key blocks, which calls for an efficient sparse attention that exploits that structure without a catastrophic loss in quality.

Sol-Attn is training-free and calibration-free. It folds block selection into the online-softmax pass itself, so there is no separate routing stage to pay for:


03 — Sol Engine

Agent-Native Auto Optimization

Three kinds of specificity separate a published method from a deployed model.

The community already has the toolbox — sparse attention, diffusion caching, kernel fusion — and every method comes with convincing numbers. It also comes with knobs. Any real deployment fixes three things at once: a specific model, specific hardware, and a specific configuration of duration, frame rate, and resolution. A setting tuned under one combination rarely holds under another, so recipes do not transfer directly and each deployment has to be tuned case by case — which is why accelerating a new model has historically taken weeks of hand-tuning rather than an afternoon.

Sol Engine closes that gap by search instead of by hand. It runs one optimization search per technique family, concurrently — kernel, sparse attention, cache — each sweeping its own knobs against the real checkpoint on the real GPUs. The results cannot simply be concatenated, because the techniques interact: caching changes which steps sparse attention has to serve. So the surviving candidates go to an integrator that searches over combinations rather than over knobs, with a human validator signing off on the output video.

Search topology
Parallel optimization search
Kernel
Fusion set, graph capture, numerical verification
Sparse attention
β, block size, dense-fallback layers and steps
Cache
Policy family, threshold, warmup, reuse cap
Integrator
Global search over compositions, not knobs
Deployed recipe
3.95× · near-lossless · 4.5 hours
Each axis is searched independently and concurrently; the integrator composes the survivors, because the techniques are not independent of one another.

The expensive part of accelerating a new model was never knowing which techniques exist — it was the serial grind of trying them one at a time. Running the axes in parallel and resolving their interactions in one integration step is what turns that grind into an afternoon. Nothing in the pipeline is specific to MiniMax-H3: published Sol Engine results already span a 64B Cosmos3-Super, a 22B LTX-2.3, and a 2B SANA-Video at 2.27×, 2.38×, and 2.77× on B200. H3 became the fourth on the day its weights appeared.


References

Citation

If this work is useful to you, please cite Sol Engine and Sol-Attn.

Sol Video Inference Engine
@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},
url = {https://arxiv.org/abs/2606.23743},
}
Sol-Attn
@misc{li2026solattnacceleratingvideogeneration,
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},
url = {https://arxiv.org/abs/2607.24027},
}