MiniMax: H3

MiniMax: H3 is a video generation model from MiniMax, released on 2026-07-29. It is described by the provider as a lightweight, open-weights model built for precise multimodal editing and controlled content generation. Its main category is video, and it is aimed at workflows where the output is a generated or edited video clip rather than text or a still image.

The model accepts text, image, video and audio as inputs, and produces video as its only output modality. Generation is controlled through several settings: output resolutions are 2K, and you can choose from aspect ratios of 21:9, 16:9, 4:3, 1:1, 3:4 and 9:16. Clip durations range from 5 to 15 seconds in one second steps. You can supply a first frame, a last frame, or both to anchor the result, and audio generation is supported. No context length is published for this model.

Those capabilities point to a fairly specific set of uses. Instruction guided edits fit well, since the model takes an existing video alongside text instructions. Text and brand rendering inside generated footage is another stated strength, which suits short promotional or title sequences. Because it accepts image, video and audio inputs together, it can be used for controlled content generation where you want to steer the look and timing of a clip rather than leave everything to a prompt. The first frame and last frame options make it practical for interpolation style work, filling the motion between two fixed images. The range of aspect ratios covers both landscape and vertical delivery, and the 5 to 15 second durations suit short form clips. Audio generation means a single call can return a clip with sound rather than requiring a separate audio step.

You call the model through api14's OpenAI compatible API using the model id "minimax/hailuo-3". The request parameters it supports are max_tokens, temperature and top_p, so sampling behaviour is controlled in the familiar way. Note that maxOutput is reported as 0, so do not rely on an output token budget as a length control for the video itself; use the duration and resolution settings instead.

There are a few cases where another model may fit better. If your task is text generation, image generation or audio only output, this model will not serve it, since video is its sole output. If you need a published context window to reason about long prompt histories, that figure is not available here. And if you need clips longer than 15 seconds, or resolutions other than 2K, the listed options do not cover that.