A 33B omni-modal video model, running on hardware that sits on a desk. The Sol Video Inference Engine delivers a 3.92× on DGX Spark and 4.52× on GeForce RTX 5090 at near-lossless quality. It gets there without distillation, without LoRA or fine-tuning, and without an offline calibration pass.
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.
Measured on NVIDIA DGX Spark and GeForce RTX 5090 · 1344×768 · 24 fps · 124 frames.
The Sol Video Inference Engine converged on a recipe that generates the same clip 4.52× faster with no perceptible change in picture or sound. It gets there without any of the machinery usually reached for when a model this size has to fit on one card: 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.
Each technique multiplies onto the one before it, and together they compose to the final end-to-end number.
Full optimization is the whole stack — lossless kernels, Sol-Attn, the cross-step cache, and a resident BF16 VAE. Each clip carries its own audio track, so the sound is part of what has to survive. Render settings differ per device and are noted above each pair.
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]
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.
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.
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.
| Fusion | What it eliminates |
|---|---|
fused RMS-AdaLN | HBM round-trips across the modulation branch — the largest single win on H3, given its 13B AdaLN share |
fused QKNorm + RoPE | Two full read/write passes over Q and K, plus 3D MM-RoPE indexing overhead |
QKV merge | Three skinny GEMMs collapsed into one wide, tensor-core-friendly GEMM |
GEMM + GELU | An activation-sized HBM round-trip per MLP block |
torch.compile capture | Residual launch overhead and leftover elementwise chains |
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.
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:
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.
If this work is useful to you, please cite Sol Engine and Sol-Attn.
@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},
}
@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},
}