Minimax H3

MiniMax H3: One Model for Generating, Referencing, and Editing Video

Most AI video tools do one trick well. You generate a clip in one place, add sound in another, and re-render from scratch every time a client asks for a small change. MiniMax H3 collapses that into a single multimodal model that understands text, images, video, and audio as one creative context — and returns a finished scene with sound already in it.

Mix any reference into one take

Feed H3 up to 12 files at once: as many as 9 images, 3 video clips, and 3 audio tracks. It reads the faces, the choreography, the camera work, and the voices across all of them simultaneously, then merges them into a single coherent shot. Reference a character photo to lock identity, point at a dance clip to borrow its motion, drop in a voice sample to set the tone — in the same prompt, not three passes.

Edit with a sentence, not a timeline

Say what to change and H3 changes only that. Swap a cat for a dog, turn a green screen into a fairytale forest, shift day to night, or hand a character new lines. Everything outside the edit stays pixel-stable, which is what makes shot-by-shot iteration practical instead of a gamble.

Output you can actually ship

• Up to 1440p at 24 FPS, 5–15 second clips
• Native two-channel stereo audio on every generation — ambience, effects, music, synced dialogue
• Aspect ratios from 21:9 cinematic to 9:16 vertical
• Accurate on-screen text, logos, packaging, and UI elements
• Real camera language: rack focus, handheld, hard cuts, match cuts, title cards

Who uses it

Content creators shipping daily shorts. Marketers testing ten ad directions before a briefing. E-commerce sellers turning flat product photos into listing videos. Game teams producing character PVs and UI walkthroughs. Filmmakers previsualizing scenes before anyone books a location.

No timeline, no keyframes, no install. Try MiniMax H3 free — free credits on signup, no card required.