MiniMax models and APIs

About MiniMax

MiniMax contributes the H3 line: three video families that start from a written shot, an approved opening frame or a set of reference images and return sound together with the picture, plus three LoRA families that run the open H3 weights with your own LoRA files.

MiniMax H3 publishes all three tasks. Its text-to-video returns 5 to 15 seconds across six aspect ratios; its reference task holds a character, a wardrobe and a location steady using up to nine reference images. MiniMax H3 Max covers text-to-video and image-to-video and can land on a second frame you supply. MiniMax H3 Spicy is the uncensored image-to-video family, animating your own still into 3 to 15 seconds at 480p, 540p, 768p or 1080p with the prompt optional. MiniMax H3 LoRA runs text-to-video, image-to-video and reference-to-video on the open-weights H3 with up to three of your own LoRAs per request, at 480p to 1080p for 3 to 15 seconds. MiniMax H3 Singularity LoRA offers image-to-video and reference-to-video on a community fine-tune of H3 built for fast action, dance and wide shots. MiniMax H3 Image LoRA uses the same weights to make single stills and edits with up to nine references, so a key frame and the clip that follows can share one model and one LoRA.

If continuity across several shots is the problem, start with the H3 reference task: nine reference images is a large budget for one call and it removes most of the work a prompt would otherwise have to do. Use H3 Max when a specific closing frame matters. Check each family page for the resolution tiers it actually publishes — they are not identical across the families. Pick a LoRA family when a character or a style already lives in weights trained for H3; the same LoRA files load on all three.