MiniMax H3 is live on Skiprr
MiniMax's new multimodal video model renders up to 2K and takes direction three ways: a prompt, a first and last frame, or a set of reference images, clips and audio. Four to fifteen seconds, with sound. It is in the model picker on every Skiprr video production today.
MiniMax is the lab behind Hailuo, and H3 is its new multimodal video model. It is in the Skiprr video model picker now. No keys, no waitlist, no separate account, and billed in the same credits as the rest of your workspace.
promptA barista in a small sunlit cafe pours steamed milk into a ceramic cup, forming a leaf in the foam. Close on the hands and the cup, steam rising, a brass espresso machine soft in the background. Warm morning light, natural colour.
We picked that shot because it is hard. Hands are where video models usually come apart, poured liquid has to obey gravity, and latte art is a pattern that has to form in the right order. The pour is continuous, the leaf builds from the base up the way a real one does, and the hand holding the cup keeps its fingers.
Three ways to direct it
Write a prompt and H3 makes the shot from nothing, in any of six aspect ratios from 21:9 to 9:16.
Give it a first frame and it animates that image. Add a last frame and you have fixed both ends of the shot, so the model is working out how to get from one to the other rather than deciding where to go.
promptThe trainers sit on the plinth as the afternoon light slides slowly across the gallery wall; dust drifts through the beam. The camera eases toward the shoes.
Or skip frames altogether and hand it references: up to nine images, three clips and three audio files, to steer character, motion, camera, style and rhythm.
One rule to know, because it is the model's and not ours. Frames and references are separate modes. H3 will not accept a first or last frame together with reference media, so choose one approach per scene. When a Skiprr scene already has a composed keyframe, the keyframe is what gets sent, and anything the references were carrying belongs in the keyframe or the prompt.
Resolution and length
H3 renders at two sizes, 768P and 2K, and 2K is the reason to choose it for anything that will be seen large. Clips run any whole number of seconds from four to fifteen, so a scene can be as long as the beat needs rather than snapped to a fixed length. They come back with sound.
How to run it in Skiprr
Pick MiniMax H3 in the model picker on a video production. Iterate at 768P and render the keeper at 2K, which costs a little over half as much again. Describe one clear action and one camera move per scene, and when you pin a last frame, say how the shot should arrive there. Seedance 2.5 stays preselected as the default. Your brand brain still runs the show, so references, voice and guardrails carry into every generation.