MiniMax H3

MiniMax H3 is a video model for creators that turns text, images, video, and audio references into short clips with synchronized sound.

A reference can direct more than the opening frame

A first-frame generator animates a picture. H3 can also read a collection of reference media and use it to direct a shot. A character photo can define who appears, a video can supply movement, and an audio clip can guide the voice or sound. The prompt explains which parts of those references belong in the result. This makes the model relevant to creators working on recurring characters, product demonstrations, and scenes whose action must follow an existing performance.

• Text mode starts from a written scene description and generates video with audio.
• First/last-frame mode uses an opening image, an ending image, or both to guide the shot's boundaries.
• Reference mode combines text with image, video, or audio inputs. In the released model this route is called Ref2VA: reference to video and audio.
• The official system describes 4-15 second output, native stereo audio, and multiple landscape, portrait, and square aspect ratios.

The original announcement is dated 31 July 2026. MiniMax published its open-weight announcement on 3 August. These are separate milestones: the hosted product arrived before the public model weights. MiniMax points creators to Hailuo's H3 tool for a browser-based experience.

What the public weights actually include

The official repository separates the system into prompt interpretation, base generation, and 2K regeneration. The distinction matters when choosing between running a model yourself and buying a complete hosted render. Downloading the base checkpoints does not reproduce every component of the hosted system.

• H3-Base FL2VA handles text and first/last-frame generation. H3-Base Ref2VA handles reference media. Both produce video and audio.
• The base model's documented output uses a 768-pixel shorter side. The full 2K workflow combines local base generation with official API components.
• H3-Context-IR prepares the structured context used by generation. H3-Regenerate-2K regenerates a lower-resolution result while retaining the original context.
• The repository says those two surrounding modules are not yet open-sourced. Its full-workflow examples call the MiniMax platform for them.
• The weights use the MiniMax H3 Community License Agreement. Read that license before deployment or redistribution; the download is not an Apache-2.0 grant.

The model download on Hugging Face includes separate FL2VA and Ref2VA folders. The repository recommends SGLang, vLLM, diffusers, and ComfyUI workflows. Its SGLang example serves a checkpoint across four GPUs, so local deployment is a developer route requiring suitable hardware rather than a lightweight desktop installation.

H3 Max and the meaning of Ref

H3 Max is fal's post-trained variant of MiniMax H3. fal describes changes aimed at prompt adherence, aesthetics, and throughput. It offers separate text-to-video, image-to-video, and reference-to-video endpoints. The phrase H3 Max Ref refers to that reference workflow, rather than a separate MiniMax product to install.

In fal's reference API, the prompt addresses assets as Image 1, Video 1, or Audio 1. For example, two character pictures can define the subjects while the prompt asks them to walk together through a garden. The API accepts up to 12 reference files in total. Video and audio references are 2-15 seconds per clip, with a combined duration limit of 15 seconds for each modality. Optional first and last frames provide additional control over the beginning and ending.

Choose the route before estimating the bill

Use Hailuo for the browser workflow, the MiniMax platform for the official hosted API, or the released base checkpoints for a self-hosted pipeline. fal's Max endpoints are another hosted route with their own input limits and billing. The reference endpoint lists output charges and a separate reference-input token charge, so the length of the finished video alone does not determine its total cost. Check the selected provider's current quote before rendering, especially when uploading video references.