MiniMax H3 API for Video Generation and Editing

The MiniMax H3 API gives developers programmatic access to MiniMax H3, an open general purpose multimodal video model that treats text, images, video, and audio as one shared context. Rather than splitting generation, editing, and reference into separate task models, H3 reads the full input set and returns a finished clip with sound. On Atlas Cloud it runs behind a single API key with pay-as-you-go pricing.

Brand films, trailers, vertical short drama, product and ecommerce spots, game and UI motion demos, and stylized animation all sit inside its range. Because the model handles on screen text, subtitles, and brand assets, teams also use it for concept validation, storyboard previews, and visual pitches before committing production budget.

Create an Atlas Cloud account, generate an API key, then send a prompt plus any reference files to the video generation endpoint and poll for the finished result. Passing media as hosted URLs is recommended over inline uploads, since the request body is capped at 64MB. Start building today.

Billing is pay-as-you-go, so you pay per call instead of buying a subscription or a seat license. Cost tracks what you actually render, which means resolution and clip length drive the total for a batch. Check the model page for the current per generation rate before planning large volume runs.

Yes. Every H3 result is delivered with sound in native stereo, so dialogue, effects, and ambience arrive in the same pass as the picture. Supply a reference audio clip and the model can carry that timbre onto a character, which removes a separate voice synthesis step from the pipeline.

Clips run from 5 to 15 seconds at 24 FPS. In 1440p mode the short side renders at 1440 pixels for ratios between 16:9 and 9:16, and outside that band the frame holds roughly 3.7M pixels in total, for example 2976 by 1248 at 21:9. Text to video and omni reference requests accept 21:9, 16:9, 4:3, 1:1, 3:4, and 9:16, while first and last frame requests inherit the aspect ratio of the input image.

Up to nine images, three video segments, and three audio clips can travel with one prompt, capped at twelve files in total. Video and audio each stay within 15 seconds combined, and audio must accompany an image or a video rather than arrive on its own. Prompts reach 7000 characters, which leaves room for shot by shot direction covering camera, performance, and sound.

Accepted inputs include H.264 and H.265 video, JPG, JPEG, PNG, WEBP, HEIC, and HEIF images, and WAV or MP3 audio, with AAC or MP3 for the audio track inside a video file. Per file ceilings are 50MB for video, 30MB for images, and 15MB for audio. Because the request body is limited to 64MB, hosted URLs remain the safer route for heavy assets.

Editing is one of its core modes. Send a source clip with instructions and the MiniMax H3 API can swap characters or objects, replace backgrounds and lighting, layer in visual effects, and rewrite dialogue while keeping untouched regions stable. Compound instructions are handled in one request, so several changes land together instead of across repeated round trips.

Most video models take one prompt plus one image and return a silent clip. H3 instead consumes a mixed set of references, reads character, motion, camera, and sound intent across all of them, then returns a clip with native audio. If your pipeline currently stitches a video model, a voice model, and an editor together, H3 collapses those stages into a single call.