MiniMax H3

MiniMax H3 is an AI video generator that takes text, images, video clips, and audio files as input and produces clips up to 15 seconds long at 2K resolution and 24fps — with stereo audio generated in the same pass as the picture.

What makes it different from other AI video tools

Most AI video generators treat audio as a separate problem. You generate the clip, then source or produce sound elsewhere. MiniMax H3 handles both in a single generation: describe the dialogue, ambience, foley, or music in the prompt, and the model produces a video with audio already timed to the shot.

The other notable difference is how it handles reference material. Rather than accepting a single image or prompt, MiniMax H3 reads up to nine images, three video clips, and three audio files in one shared context. A character photo can establish identity, a video clip can define camera movement, and an audio recording can guide voice or rhythm — all feeding into a single generation rather than separate tasks.

Generation workflows

MiniMax H3 supports four distinct ways to create or modify video.

Text to video is the starting point for most users. A well-structured prompt can carry subject description, camera movement, lighting, visual style, sound design, and constraints — producing a complete clip from a written brief alone.

First and last frame lets you upload one or two keyframe images to define how a scene opens and closes. The model generates the motion between those frames. This is particularly useful for product reveals, scene transitions, and storyboard animation where the start and end states need to match specific visuals.

Reference to video combines images, clips, and audio into one generation context. Each uploaded asset can be identified by its order in the prompt, giving more precise control over how individual references shape the result. It is suited to work where consistency matters — matching a product across shots, preserving a character's appearance, or replicating a camera style from existing footage.

Instruction-based editing works on footage you already have. Upload a clip, describe what should change and what must stay the same, and MiniMax H3 applies targeted edits — swapping a product, updating on-screen text, shifting the time of day — without rebuilding the whole shot.

Who uses MiniMax H3

The tool sits at the intersection of advertising, social content, and production pre-visualization. Creative directors use it to prototype ad concepts with synced dialogue before committing to a full shoot. Motion editors use it to generate footage that is detailed enough to cut into a timeline without upscaling. Social media creators use the seven supported aspect ratios — from 21:9 cinematic down to 9:16 vertical — to produce platform-specific content from a single workflow.

Game developers and cinematics teams use it to block out camera direction and scene timing early in production. Product marketers use the reference and editing modes to generate and iterate on e-commerce visuals without repeated reshoots.

Output specs and limits

Clips run from 5 to 15 seconds at up to 2K (1440px short edge) and 24fps. The 15-second ceiling is the main constraint for longer-form work; there is no way to generate a continuous clip beyond that length in a single pass. Text and brand elements in generated frames are identified as a supported capability but require manual review before use in published work.

An open-weight release has been announced by MiniMax, which would allow self-hosted and customized deployments, but it is not yet available.

Pricing and access

MiniMax H3 uses a one-time credit model with no subscription or auto-renewal. Three packs are available: Starter at $9.90 (200 credits), Creator at $29.90 (650 credits), and Studio at $99.90 (2,500 credits). Each 5-second generation costs 25 credits, putting the per-video price between $1.00 and $1.24 depending on pack size. Credits never expire, and failed or rejected generations are refunded automatically.

The first generation is free with no credit card required, which makes it straightforward to test prompt structure and reference handling before purchasing credits.