MiniMax-H3 API for 2K Multimodal Video Generation

MiniMax H3 is MiniMax's general-purpose, open-weight video generation system. Instead of treating text-to-video, keyframe animation, and multimodal references as unrelated tools, it interprets text, images, video, and audio within one context. The hosted API uses an asynchronous workflow: submit a task, store its task ID, check its status, and retrieve the finished video URL.

Output runs from 4 to 15 seconds at 24 fps, with 768P and 2K modes, landscape, square, and portrait ratios, and native stereo audio. The released H3-Base weights use the MiniMax H3 Community License; hosted API access and self-hosting remain separate deployment choices.

Which MiniMax H3 Input Mode Should You Use?

Choose the input mode before building the request. First/last-frame inputs and multimodal reference inputs are separate modes in the official API and cannot be mixed in one generation task.

For reference-to-video, MiniMax H3 accepts up to nine images, three video clips, and three audio clips, with a maximum of 12 files in total. Each reference video or audio clip can be 2–15 seconds, and the total duration for each media type cannot exceed 15 seconds. Use public URLs for large assets because the complete request body is limited to 64 MB.

MiniMax H3 API for Ecommerce and Batch Video Workflows

For ecommerce, supply approved product images, state which details must remain unchanged, and add a reference video only when its motion or camera path matters. First-and-last-frame mode is often simpler for controlled product reveals.

For batch generation, treat every variation as a separate asynchronous job. Store its prompt, asset URLs, settings, task ID, and status; use a bounded queue; make retries idempotent; and copy successful outputs to your own storage. Keep logo treatment, framing, lighting, camera language, and CTA fixed while injecting variable SKU data per job.

MiniMax H3 vs Seedance 2.5 vs Veo 3.1

These models overlap but solve different constraints. MiniMax H3 emphasizes 2K output, open weights, stereo audio, and mixed references. Seedance 2.5 supports longer generations and larger reference sets. Veo 3.1 fits short Google Cloud workflows with a documented 4K path.

Compare duration, reference capacity, audio, retries, and finished-clip cost—not maximum resolution alone. Seedance 2.5 may reduce scene stitching; H3 may simplify mixed-reference 2K work; Veo 3.1 may suit short Google-centered 4K workflows.

How to Write Better MiniMax H3 Prompts

Write the prompt as a compact production brief that separates source roles, timeline, camera direction, audio, preserved details, and exclusions.

Prompt formula: subject and reference roles + scene goal + timed actions + camera movement + lighting and visual style + dialogue, sound effects, and music + details to preserve + final frame

Ecommerce Product Video Prompt

Use Image 1 for product identity and Image 2 for lighting. Create a 10-second vertical launch video for the matte-black speaker. Begin with a macro grille shot, pull back as it rotates, then show water droplets vibrating with the bass. Preserve the logo, controls, proportions, and finish. Use cool edge light, restrained motion, low electronic ambience, and no extra text.

First-and-Last-Frame Transition Prompt

Start from the supplied closed package and end exactly on the supplied final frame with the product assembled. Use one continuous camera move: the box opens, components rise, and the product locks into place. Preserve package graphics and final geometry. Add mechanical clicks and a short resolved tone.