MiniMax H3 API Guide: 2K Video, Native Audio & Limits

MiniMax officially launched MiniMax H3 on July 31, 2026. MiniMax positions it as a general-purpose omni-modal generation model. The public developer workflow currently focuses on video generation, with required 2K output, native stereo sound, and durations from 4 to 15 seconds.

This guide separates the available API from the features MiniMax says are still coming. MiniMax's launch post describes video at up to 2K and 15 seconds; the launch-day API is more specific, requiring 2K and accepting whole-number durations from 4 through 15.

What is MiniMax H3?

MiniMax describes H3 as an omni-modal generation model. In practical terms, you can give it several kinds of reference material and explain how they relate in plain language. A prompt might ask H3 to borrow camera movement from a video, preserve a character from an image, and follow the timing or voice from an audio reference.

That is different from a workflow made of separate motion, lip-sync, sound-effect, and upscaling tools. H3 is designed to understand the combined context before generating the result. Its audio is produced as native stereo sound rather than added as a separate post-production track.

The official launch post also describes image generation, editing, audio generation, and audio-video editing as parts of H3's broader training design. The public H3 API documented at launch, however, is a video-generation endpoint. Treat those two statements separately: H3's research scope is broad, while the currently documented developer workflow focuses on video.

What the MiniMax H3 API supports today

The official API uses the model name MiniMax-H3 and exposes three main workflows:

1. Text to video. Send a required text prompt and choose a supported aspect ratio.
2. Image to video. Add a first frame, a last frame, or both. The input image determines the output ratio.
3. Reference to video. Combine reference images, videos, and audio, then use the prompt to explain what H3 should take from each one.

Current output settings are specific:

• Resolution: 2K
• Duration: any whole number from 4 to 15 seconds
• Aspect ratios: adaptive, 21:9, 16:9, 4:3, 1:1, 3:4, or 9:16. Text-to-video requires a concrete ratio; image-to-video always follows the input image; reference-to-video can default to adaptive.
• Reference images: up to 9
• Reference videos: up to 3, with a combined duration of no more than 15 seconds
• Reference audio files: up to 3, with a combined duration of no more than 15 seconds
• Total request body: no more than 64 MB

Every request needs a non-empty text prompt. Audio cannot be the only reference: a reference-to-video request must include at least one image or video. The API also keeps image-to-video and reference-to-video as separate modes, so first/last-frame inputs cannot be mixed with reference inputs in one request.

Those details matter. A prompt can be excellent and still fail if its media roles or file limits do not match the endpoint.

Create a task by sending POST to /v2/video_generation on api.minimax.io, using Bearer authentication and a content array that includes at least one non-empty text item. The response returns a task_id. Poll GET /v2/query/video_generation/ with that task ID until the task succeeds, then read the video URL from task.content.url.

Why native audio and mixed references matter

The interesting part of H3 is not simply “video with sound.” It is the relationship between the inputs.

For a product clip, one image can establish the product and packaging, a short video can establish the camera move, and an audio clip can establish rhythm. For a character scene, an image can anchor appearance while a video provides motion and an audio reference guides the performance. The text prompt connects those sources by saying what to preserve and what to change.

A useful prompt therefore describes relationships, not just atmosphere. Name each reference, assign it one job, and state the intended output in time order.

For example:

Reference Image 1:
keep the face, clothing, and color palette.
Reference Video 1:
follow the slow push-in camera movement.
Reference Audio 1:
use its vocal tone and timing.
The character turns toward the window,
says one short line, then pauses.
City lights reflect across the glass.
Keep identity and room layout consistent.
Native stereo ambience. No background music.

This structure is easier to revise than a long paragraph where every reference is asked to control everything.

A practical three-step workflow

1. Choose one generation mode

Start with the result you need. Use text-to-video for a scene with no visual anchor. Use first/last-frame image-to-video when composition matters. Use reference-to-video when identity, motion, or audio must come from existing media.

Do not add references just because the API accepts them. Each file should solve a specific continuity or control problem.

2. Write the prompt as a shot plan

Describe the subject, action, camera, lighting, sound, and ending. When you use references, say exactly what each one contributes. For a 4–15 second clip, one clear dramatic beat is usually more controllable than a crowded sequence of unrelated events.

You can draft the shot in our text-to-video workspace or plan motion in the image-to-video workspace. These third-party workspaces are not the MiniMax H3 API. Check the model and settings shown in the interface before treating an output as an H3 result.

3. Generate asynchronously and review the result

The H3 API is asynchronous. Creating a task returns a task_id; a second endpoint reports whether the job is queued, running, succeeded, failed, or expired. When it succeeds, the response includes the video URL.

Review more than visual polish. Check identity, object placement, camera intent, spoken words, sound timing, and whether the ending gives you a usable edit point. Change one variable at a time on the next attempt.

The technology behind H3, in plain language

MiniMax names four core systems in the launch post:

• Contextual Omni Representation describes the relationships among text, image, video, and audio so the model can follow cross-modal instructions.
• H3-VAE compresses media efficiently. MiniMax says it raises effective sequence length to four times its previous level and helps make native 2K output practical.
• H3-Omni Transformer separates understanding and generation workloads while training. MiniMax reports that this improved end-to-end training throughput by nearly 30%.
• In-Context Regeneration lets the base model regenerate a lower-resolution result at 2K while looking at the original references again, instead of relying only on a conventional super-resolution step.

For 2K output, MiniMax says the regeneration stage consults the original multimodal context again, helping recover details that a conventional super-resolution pass would have to infer.

Limits to know before production

H3 is available, but the launch documentation still sets clear boundaries.

The API currently documents 2K output only, durations are limited to 4–15 seconds, and reference inputs must follow strict role and size rules. MiniMax also says visual detail can still improve in some scenarios. Its official post says model weights are planned for release in the coming days, subject to applicable laws and regulations; planned open weights are not the same as weights already being available.

For paid or client work, run a small test with the exact aspect ratio, reference mix, and audio requirements you intend to use. Keep the successful prompt, input roles, and output settings together so the shot can be reproduced.

The bottom line

The MiniMax-H3 API is available for 2K video generation with text, image, video, and audio context. At launch, MiniMax said the model weights were still planned for release in the following days.

A practical starting point is to give each reference one clear role, test one short beat, and verify the current API limits before expanding the shot or scaling a production workflow.

MiniMax H3 (minimax3.com) is an independent third-party platform and is not affiliated with, endorsed by, or operated by MiniMax or Hailuo. See the official MiniMax H3 announcement and official API documentation for the latest provider information.