Model brief
MiniMax H3
MiniMax H3 is a general-purpose multimodal generation model that produces video with native stereo audio, up to fifteen seconds at 2K, from mixed text, image, video and audio context. MiniMax announced it on 31 July 2026 with the intent to release open weights; fal's later H3 Max announcement refers to it as an open-weights model.
No tracked platform states that it runs on this model directly, so this brief is kept out of search indexing. It is recorded because it is the base of a model that is deployed.
- Developer
- MiniMax
- Type
- Multimodal generation (video, image, audio)
- Released
- 31 Jul 2026
- Native audio
- Yes
- Real-time capable
- Not publicly disclosed
- Open weights
- Yes
- Public API
- Not publicly disclosed
- Based on
- —
- Tracked deployments
- 0
- Last verified
In this directory H3 appears mainly as the base model behind fal's H3 Max, which is what the live channels we track actually run on.
Platforms running on MiniMax H3
No tracked platform currently states that it runs on this model.
Timeline
- · Model Change
fal announced H3 Max, a post-trained variant of MiniMax H3 that generates a five-second clip in about three seconds. It is the model the live channels we currently track are built on.
- · Research Release
MiniMax introduced H3, a multimodal generation model producing up to fifteen seconds of 2K video with native stereo sound, and said it would open the weights.
Related models and reading
The phrase gets used for everything from a caption bot to a channel that paints itself into existence. This guide draws the line, explains how the real thing works, and points to channels you can watch right now.
Official sources
- MiniMax H3: An Open Model Breaking the Boundaries Between Tasks and Modalities
“generating video with native stereo sound, up to 15 seconds at 2K resolution”
Launch post dated 31 Jul 2026. Describes H3 as a general-purpose multimodal generation model with native stereo audio and states the intent to release open weights. fal's 27 Aug post later refers to H3 as open-weights, so the open-source flag is recorded as yes with both records attached.
nameofficial_urldeveloper_organization_idmodel_typerelease_datesupports_native_audiois_open_source