MiniMax H3 Opens Its Video Weights, but Not Where Many Developers Work

MiniMax H3 promises a unified route from prompts and reference media to video with synchronized stereo sound. The catch is that its downloadable weights are neither globally usable nor equivalent to the full product: the default license excludes the US, EU, UK and South Korea, while 2K generation remains tied to the hosted API.

One model, mixed inputs

Released July 31, H3 accepts text, images, video and audio as a single input context. It generates video and synchronized 32 kHz stereo audio in one pass; hosted clips run from four to 15 seconds at 24 frames per second and can reach 2K resolution.

That unified input is the practical distinction H3 is trying to make. Rather than requiring separate tools for video generation, voice and sound effects, a user can supply reference media and direct the result in natural language. Its omni-reference mode accepts as many as 12 mixed files; a separate variant supports zero, one or two input images, including first- and last-frame control.

The package is technically substantial. H3’s core single-stream transformer has 33 billion parameters, while the full inference stack, including its text encoder, video encoder-decoder and multimodal-input system, totals about 69.2 billion parameters. MiniMax says its video autoencoder compresses the sequence representation, allowing longer or higher-resolution generation without a proportional increase in compute.

The open-weight version is a narrower system

MiniMax published the weights on Hugging Face on August 3 under its Community License Agreement. In the excluded territories, that agreement does not authorize local use, modification, distribution, deployment or use of locally generated outputs. The hosted service is available there, drawing a firm distinction between using H3 and self-hosting it.

The missing components explain why downloading the weights does not recreate the product MiniMax sells through its API. The excluded Context-IR system interprets mixed inputs, while H3-Regenerate-2K produces higher-resolution video by regenerating the model’s lower-resolution output in context. The API charges an additional $0.05 per output second for that 768p-to-2K pipeline.

Commercial terms for permitted users

• Organizations with annual revenue below $20 million can use H3 commercially without charge under the community license.
• Organizations above that threshold need MiniMax’s prior written authorization.
• The license also bars users from training or distilling other AI models with H3 outputs.

Self-hosting is also a meaningful hardware commitment, even where the license permits it. One task partition requires about 134 GiB of storage; running both primary partitions requires roughly 196 GiB after deduplication. The reported consumer setup uses two RTX 5090 GPUs plus 384 GB of system memory and takes about 9.3 minutes to generate a five-second clip, versus roughly 13 seconds on four H100 GPUs.

Strong rankings, with a different buying decision

H3 ranked third in Artificial Analysis’s audio-enabled text-to-video leaderboard and second in image-to-video with audio as of August, according to the article’s account of those results. In its open-weights text-to-video ranking without audio, the model placed first with an Elo score of 1,303. Those are useful signals of relative output preference, but they do not eliminate the deployment distinction between the downloadable base model and the hosted stack.

MiniMax lists hosted pricing at $7.80 per minute for 2K output and $4.80 for 768p. The source’s comparison puts the 2K rate below Veo 3.1’s listed $24 per minute for 1080p, though resolution alone does not settle output quality or workflow fit. For teams evaluating H3, performance and price are only part of the decision; location, local-compute capacity and access to the API-only components determine which version they can actually use.