01
Turbo and Lightning LoRAs
4-step and 8-step LoRAs cut the step count. ComfyUI uses 8 steps for text and image to video and 4 for reference to video.
Blog · October 9, 2026
MiniMax H3 inference speed: steps, generation time and how to make it faster
Base MiniMax H3 uses many denoising steps, so one 5-second 768p clip takes minutes on a single GPU. Step distillation, sparse attention and a tuned serving stack bring that down to seconds. This page collects the published numbers and gives the source for each.
Steps
How many steps does MiniMax H3 use?
| Setup | Steps | Notes |
|---|---|---|
| MiniMax model card and README | Not stated | The weights are CFG-distilled: no guidance scale and no negative prompt |
| ComfyUI template | 20 | res_multistep sampler, simple scheduler. Simple shots hold up at 12–16 steps |
| SGLang cookbook | 50 | 50 sigma points = 49 model calls |
| ComfyUI Lightning LoRA | 8 or 4 | 8 for text and image to video, 4 for reference to video |
| FastH3 V2 | 8 model calls | Distilled with DMD2, 80% sparse video attention |
Because the weights are CFG-distilled, each step is one model call. Fewer steps means fewer model calls, and the model calls take most of the time.
Sources: MiniMax H3 model card, ComfyUI docs, SGLang cookbook, FastH3 V2 model card. Read October 9, 2026.
Generation time
MiniMax H3 generation time by GPU
| GPUs | Clip | Setup | Time |
|---|---|---|---|
| 1× B200 | 1344×768, 5 s / 10 s / 15 s | Base H3, 50 steps | 132.5 s / 377.4 s / 678.7 s |
| 4× H100 80GB | 1344×768, 5.17 s | Base H3, 50 steps, SGLang | 200.1 s |
| 4× H200 | 1344×768 | Base H3, 50 steps, SGLang | 74.4 s |
| 4× GB300 | 768p, 5 s | Base H3, 50 steps, SGLang | 33.1 s |
| 8× B300 | 10 s | Base H3, 50 steps, vLLM-Omni | 56.9 s |
| 2× RTX 5090 | 1344×768, 124 frames | Base H3, 50 steps, layer offload | 559.7 s |
Sources: Hao AI Lab (B200), Hyperstack (H100), SGLang cookbook (H200, GB300, RTX 5090), vLLM (B300). Read October 9, 2026; GB300 read October 28, 2026.
Times change with the tool, resolution and clip length. Compare two rows only when the setup is the same.
FastH3
FastH3: 8 model calls instead of 49
| GPU | 5 s clip, 832×480 | 5 s clip, 1344×768 |
|---|---|---|
| RTX PRO 6000 96GB | 13.5 s | 36.5 s |
| RTX 5090 32GB | 14.8 s | 38.6 s |
| RTX 4090 24GB | 54.6 s | 154.6 s |
FastH3 V2, end to end with audio, measured by Hao AI Lab (October 6, 2026). On one B200, FastH3 V1 renders a 5-second clip in 16.2 s against 132.5 s for base H3, and a 15-second clip in 47.2 s against 678.7 s (source). On eight B300 GPUs with vLLM-Omni, a 10-second clip takes about 8.7 s (source). On four GB300 GPUs with SGLang, FastH3 V2 renders a 5-second 1344×768 clip in 5.2 s and a 15-second clip in 18.2 s (median, 1,000-token prompts; source, read October 28, 2026).
FastH3-Omni adds image, video and audio references. On eight GB200 GPUs, a 5-second 768p clip from one reference photo takes 5.7 s. At 480p on local hardware, the same request takes 61.9 s on an RTX 5090 and 192 s on one DGX Spark. One reference photo makes a request 3.3–4.4× slower than text to video (Hao AI Lab, October 21, 2026; read October 23, 2026). Local times by GPU.
A Nuva Lab host with eight RTX PRO 6000 GPUs renders about 4,000 ten-second 768p clips with audio per day at full use. B200, B300 and GB200 hosts are available on request. FastH3 details.
FastH3 covers text to video with audio. Its model card says hard motion, fine detail and some audio can be below base H3. Test it on your own briefs.
Other methods
Other ways to speed up MiniMax H3
02
FP8 weights
vLLM-Omni online FP8 cut peak GPU memory by 38.9% on eight B300 GPUs. The stage time fell by 5.3%.
03
Sparse attention
The MiniMax release has full attention only. FastH3 uses 80–90% sparse video attention. vLLM measured 9.8 s dense against 7.3 s sparse for a 10-second FastH3 clip.
04
More GPUs
SGLang and vLLM-Omni split one clip across GPUs with sequence parallelism. The GPU count must divide the 56 attention heads: 1, 2, 4, 7 or 8.
In production
Production speed is a serving problem
Time per clip is one number. Production also needs batches, queues, retries and steady throughput across a whole day.
Serve FastH3-Omni in production
For servers, the FastH3-Omni model card points to UniServe, the FastVideo serving engine, with the BF16 checkpoint. UniServe has validated FastH3-Omni settings for one, two, four and eight GB200 GPUs, eight H200 GPUs and eight RTX PRO 6000 GPUs, for 768p clips of up to 15 s. The eight RTX PRO 6000 setting uses the same GPUs as the Nuva Lab standard host. One replica across every GPU gives the lowest latency. More replicas serve more requests at once.
On one replica of eight GB200 GPUs, a 5-second 768p FastH3-Omni request takes 3.9–8.7 s, and a 15-second request takes 10.5–11.9 s. The time depends on the number and type of references (UniServe evaluation, mean of 3 requests; read October 29, 2026). vLLM-Omni and SGLang also serve FastH3-Omni: see the serving notes in the requirements guide.
On a dedicated MiniMax H3 deployment, every GPU serves only your jobs, so a batch does not wait behind other customers. What the host costs.
FAQ
Questions
How many steps does MiniMax H3 use?
MiniMax does not state a default. The ComfyUI template uses 20 steps and the SGLang cookbook uses 50. FastH3 needs 8 model calls.
How long does MiniMax H3 take to generate a video?
Base H3 takes about 132 seconds for a 5-second 1344×768 clip on one B200 GPU, and about 200 seconds on four H100 GPUs at 50 steps. FastH3 V1 takes 16.2 seconds for the same clip on one B200.
How do I make MiniMax H3 faster?
Use fewer steps with FastH3 or a Turbo LoRA, use FP8 weights, use sparse attention, or split the clip across more GPUs. Test each method on your own prompts, because each one can change the output.
Is FastH3 as good as base MiniMax H3?
It depends on the shot. The FastH3 model card says hard motion, fine detail and some audio can be below base H3. Test it on your own briefs.
How do I run FastH3-Omni in production?
Serve the BF16 checkpoint with UniServe's reference settings, as the FastH3-Omni model card recommends. UniServe validated them on one to eight GB200 GPUs, eight H200 GPUs and eight RTX PRO 6000 GPUs. vLLM-Omni and SGLang also serve FastH3-Omni. Nuva Lab offers FastH3-Omni on dedicated hosts.
Sources and further reading
Read next
Get this speed on your own host
Tell us your volume and clip formats. We'll size a dedicated deployment.
Get started · 1 min