MiniMax H3 takes text, images, video clips, and audio into one shared context and returns 4-15 second clips at 24 FPS with native stereo sound - generation, referencing, and editing in a single model. This guide collects its real specifications, the documented prompt-writing rules, and every legitimate way to run it, then lets you try the same prompt style on AI Fruit's own video models in the generator below.
H3 is available through MiniMax's own app and API, through hosted providers, and as downloadable weights. Pick the route that matches how much control you need.
MiniMax ships H3 as a web app at hailuoai.video and a developer API at platform.minimax.io. This route gives you the full official pipeline, including the hosted context processing that turns free-form prompts and references into generator-ready instructions.
fal hosts H3 as a launch partner across three endpoints - Text to Video, Image to Video, and Reference to Video - published under the hailuo-03 API name. fal lists $0.26 per second of 2K output with no subscription, so short tests stay cheap.
The weights live at MiniMaxAI/MiniMax-H3 under the MiniMax H3 Community License, with a Diffusers pipeline for local inference. The open release covers the base generator and the 2K regeneration pass; the hosted context-processing system is not part of it, so local runs follow the published prompting guidance instead.
H3 unifies jobs that used to need a relay of separate models, and it ships weights you can download - a combination few video models offer.
Text, images, video, and audio enter one shared context. Character identity from a photo, camera language from a clip, a voice from a recording - H3 reads them together instead of handing off between specialists.
Every generation carries its own soundtrack: ambience, foley, and speech generated alongside the picture, with stable dialogue support for 11 languages including English and Chinese.
H3 renders legible type, which puts title cards, signage, and brand marks inside a prompt's reach. MiniMax's launch write-up calls out accurate text and brand rendering as an early-testing strength.
The complete weights are published on Hugging Face to support further development, including fine-tuning. MiniMax also positions 2K output at less than a third of mainstream per-second pricing; on fal the 2K rate is $0.26 per second.
The workflow below is documentation-based, following MiniMax's model card and published prompting guidance - pick the input mode, assemble labeled references, write the brief, then scale resolution.
No media is text-to-video. One or two frame images (opening or closing frame) is the FL2VA mode. Identity, style, motion, voice, or editing references use the Ref2VA mode.
Ref2VA accepts up to 9 images, 3 video clips of 2-15 seconds each, and 3 audio clips, capped at 12 files. Cite each one as Image 1, Video 1, Audio 1 - input order is semantic, and audio must travel with at least one image or video.
Structure the prompt in timed blocks with one primary change per beat, an explicit camera move, a directed audio track, and stated negatives. Then render at 768p and run the 2K regeneration pass for the final master.
H3 follows direction better than description. These four habits come straight from the published prompting guidance and separate usable clips from mush.
Never send unlabeled media. Write "Use Image 1 as the locked identity, Video 1 for the camera move, Audio 1 as the voice" - named jobs bind references to roles; unlabeled inputs leave the model guessing.
Mark beats as [0 to 4 seconds], [4 to 8 seconds], and so on. Give each beat one primary change and an observable end state, and put the most important beat in the middle of the timeline - the final beat is the most likely to get squeezed.
Name room tone, foley, and music placement explicitly. Tag speakers (S1, S2), put verbatim dialogue in a [Language] tag from the 11 supported languages, and describe when lips close.
"No soft dissolves, no camera shake, wardrobe identical in every shot" - negative direction lands unusually well. Name the invariant details (hair, closures, footwear) and describe transitions as physical events.
Straight answers about H3's weights, outputs, inputs, pricing, and how it relates to AI Fruit.
H3 itself runs on MiniMax's app, fal, or its open weights - but the prompt style works anywhere. Open the generator below, paste an H3-style brief, and generate with an AI Fruit model you can pick and price before spending credits.
Open the GeneratorCompare other AI video models available in AI Fruit.
Use the Seedance 2.0 AI Video Generator on AI Fruit for social fruit videos, energetic product clips, and short-form experiments with optional audio and last-frame control. The editor opens with Seedance selected so you can adjust duration, resolution, aspect ratio, and credits before generation.
View model
Use the Veo 3.1 AI Video Generator on AI Fruit for HD fruit videos, polished product shots, and short commercial scenes where native audio output and last-frame control can matter. The editor opens with Veo selected so you can review settings and credits before generation.
View model
Use the Kling 3.0 AI Video Generator on AI Fruit when fruit videos need stronger cinematic framing, optional audio, and last-frame control. The model page opens the full editor with Kling selected so creators can shape prompt, image, duration, resolution, and aspect ratio before spending credits.
View model