On September 4, NVIDIA published audited benchmarks for the Sol-H3 engine: a 5-second MiniMax-H3 video at 1344x768 resolution and 24 FPS is generated in a median of 1.653 seconds on a single 8-GPU NVIDIA B300 system — this is 4 DiT passes instead of 50 in the base model and up to 15.05x latency speedup. At the same time, the community released a set of working tools for H3 in ComfyUI: character replacement in video, accelerated face-refine for group scenes, a multi-LoRA stack loader, and a 6-step turbo merge.

image
image

What happened

NVIDIA Research released the Sol-H3 material ("Speed-of-Light MiniMax-H3 on an 8× NVIDIA B300 Blackwell System") with an audited inference benchmark: a 5-second T2VA MiniMax-H3 clip with 1344x768 parameters and 24 FPS fits into a median of 1.653 seconds on a single 8-GPU NVIDIA B300 node. The savings are gathered from several techniques at once: sparse attention with dynamic, query-dependent thresholds, fusing RMSNorm and SwiGLU kernels, INT8 quantization of QKV, FP8 transport and all-to-all Ulysses; the number of DiT passes is compressed from 50 to 4. In the same wave came community releases: ComfyUI-Viggle-Animate-H3 nodes for character replacement in video — this is a fine-tune of the H3 ref2va transformer with 33.1B parameters with DMD2 distillation and a working range of 4-8 steps; ComfyUI-H3-FaceRefine-Accelerated for group scenes (up to 9 faces in the atlas, benchmark 995.5 seconds vs 202.1 seconds, 4.925x on RTX 4090); a multi-LoRA stack loader that fixes AdaLN and quantized matching errors, which broke LoRA for H3; a 6-step flat merge of turbo LoRA (lightxv, larryvrh plus JonXL photorealism); a Windows controller that assembles H3 prompts from a plot and 1-9 references for batch generation.

Context

MiniMax H3 is a large video model that generates a clip immediately with native audio, and until recently its application was limited by two things: inference cost and a thin layer of tools. Video-DiT by design requires many sequential steps on a large transformer, so any reduction in the number of passes or the volume of attention computations is directly converted into seconds and money. Sol-H3 is interesting because it breaks down the savings by layers: less dense attention, fewer memory accesses due to fusing, less traffic between cards due to quantization, and fewer steps due to DMD2 distillation — without replacing the model architecture. The second half of the event relates to the ecosystem: the community around H3 was previously in the "downloaded the model" stage, and the listed releases close exactly those places where the workflow broke down in practice — character replacement, multi-LoRA stability, and face processing speed.

Why this matters for the industry

For the industry, this is an audited cost benchmark: the formula "5 seconds of video with audio in less than 2 seconds on a single 8-GPU node" becomes a reference point against which other video stacks will be measured. The key here is not the number itself, but its origin: the speedup is assembled from portable techniques — sparse attention, quantized transport, kernel fusing, distillation to 4-8 steps — meaning this is a reproducible engineering recipe for other video-DiTs, not a unique property of one model. For teams with 8× B300 nodes, interactive experiments and prototypes become much cheaper, and for startups this is a signal that the cost of a clip will soon be calculated the same way as the cost of tokens in text. A separate signal is the maturity of the ecosystem: around H3 a layer of working pipelines has already formed (character replacement, batch by scenario, stable multi-LoRA), and it is this that turns an open model into a product; in the coming months, expect reproductions of Sol-H3 and the transfer of these optimizations to other video models.

Why this matters for users

For those who are already running MiniMax H3 in ComfyUI, specific steps are available: clone the Viggle-Animate nodes (the int8 version weighs 21 GB, 6 steps are considered the optimum for speed and quality), install the accelerated FaceRefine for group scenes, connect the multi-LoRA stack loader if LoRA fall with AdaLN or quantized matching errors, take the 6-step flat merge of turbo LoRA for fast generation, and tie everything together with a Windows controller that assembles prompts from a plot and references. Notably, the face-refine speedup is measured on an RTX 4090 — such optimizations reach consumer cards, not just data center B300. For those who haven't touched H3, the story is useful as a clear breakdown of speed mechanics: 15x on one model is made up of attention optimizations and quantization, and the Sol-H3 demo can be viewed on the NVIDIA Research page.

What is still unknown / limitations

The 15.05x figure is the latency of one configuration (a 5-second clip 1344x768 at 24 FPS) on a single 8-GPU B300 node, not a universal speedup of MiniMax H3: on different hardware, a different duration or resolution, the result will be different. There is a minor discrepancy in the sources themselves: the channel text says up to 15.5x, the audited benchmark says 15.05x; the order of speedup matches, but the exact value should be taken from the NVIDIA material. This is a demonstration on flagship Blackwell hardware, independent reproductions of Sol-H3 and benchmarks on other GPUs are not yet available. Community tools also require caveats: the "optimum 6 steps" for Viggle-Animate is a community observation, not an official recommendation, and turbo merges and fine-tunes can change the behavior of the base model, so quality should be checked on your own tasks.

Sources

Author

Look at AI, editorial team