// ENGINE 09 — FASTVIDEO
FastVideo-FastH3-4-Step-Preview-V1-VSA-DataFree
Four-step distill of MiniMax H3, with synced audio. Open weights from Hao AI Lab: 49 sampling steps compressed into four, at preview quality.
- CREDITS
- 40 cr
- QUALITY
- great
- SPEED
- fastest
- DURATIONS
- 5 / 10 / 15s
- ASPECTS
- 16:9 · 9:16 · 1:1 · auto
- IMAGE INPUT
- text only
GENERATE
— listed here, not yet connected to the studioNOT YET RENDERABLE
FastH3 is open weights, not a hosted API — there is no inference provider serving it, and the checkpoint needs FastVideo's VSA-H3 attention backend running on four B200 GPUs by default (the GPU count has to divide H3's 56 attention heads). So this page documents the engine, but the studio cannot render with it yet. If you want to run it, the weights, the paper and the inference code are all public.
PROMPT SAMPLES
— written for this engine — use one as a starting pointPROMPT_01
Rain hammers the tin roof of a roadside diner at night; neon reflects in the wet asphalt and a single car passes, tyres hissing through the water.
PROMPT_02
A blacksmith draws glowing steel from the forge and strikes it three times; sparks scatter, the hammer rings, and the metal dims from orange to grey.
PROMPT_03
Waves break over a harbour wall in heavy weather, spray bursting upward as gulls cry and the wind buffets the microphone.
PROMPT_04
A train pulls into an empty rural station at dusk; brakes squeal, doors hiss open, and nobody steps off as crickets start up in the grass.
PROMPT_05
Coffee pours into a glass cup in close-up, steam curling upward, the pour ringing higher as the cup fills.
PROMPT_06
A thunderstorm crosses a wheat field in timelapse: the light drops, the crop bends in the gust, and the first heavy drops hit the dust.
ABOUT
FastH3 4-Step is Hao AI Lab's distillation of MiniMax H3, the 35B dual-modality model that generates video and its audio track together. Where the base model walks 49 denoising steps, this checkpoint reaches a frame in four transformer forwards — the same pipeline, roughly a twelfth of the compute.
The distillation is data-free DMD2 paired with VSA-H3 sparse attention at 90% sparsity, meaning it was trained without a dataset of target videos: the student learns to match the teacher's output distribution directly. FastVideo report up to a 14x speedup on a single Blackwell GPU — a 15-second clip in 47.2 seconds on one B200, or 12.88 seconds spread across eight.
It renders 1344x768 at 24fps in lengths of 5, 10 or 15 seconds, with stereo audio generated alongside the picture rather than dubbed on afterwards. Weights are public under the MiniMax H3 Community License.
Two caveats worth stating plainly. This is a preview, and FastVideo say difficult motion, fine detail and some audio still fall short of base H3 — four steps buys speed, and the bill is paid in the hardest frames. And only the text-to-video-audio path was distilled: the first-last-frame and reference-to-video modes of the base model were not, so this is a text-prompt engine only.
STRENGTHS
— what this engine does better than the others- Four sampling steps instead of 49 — roughly 12x fewer transformer evaluations
- Up to 14x faster than base MiniMax H3 on a single Blackwell GPU
- Video and stereo audio generated together in one pass, natively in sync
- 1344x768 at 24fps, in 5, 10 or 15-second lengths
- Open weights under the MiniMax H3 Community License — self-hostable
PROMPTING TIPS
— how to talk to this model- TIP_01Write for a text-only engine. The first-last-frame and reference modes were not distilled, so an uploaded image has nothing to attach to here.
- TIP_02Name the sound you want. Audio is generated with the picture, so ambience and effects belong in the prompt rather than in a later pass.
- TIP_03Keep motion simple in a preview checkpoint. Four steps handles steady action well and struggles with fast, complex movement — that is where distillation shows.
- TIP_04Treat detail as the trade. If a shot depends on fine texture or small faces holding up, base MiniMax H3 at full steps is still the stronger renderer.
USE CASES
01
Fast iteration
Four steps make the try-and-look-again loop quick enough to explore a shot properly before committing to a slower engine.
02
Sound-led shots
Scenes where the audio matters as much as the picture — rain, machinery, a room tone — generated in sync rather than added later.
03
Self-hosted pipelines
Open weights mean it can run inside your own infrastructure, which matters when the footage or the prompts cannot leave the building.
04
Distillation research
A published reference point for what data-free DMD2 and sparse attention cost in quality at four steps.