MiniMax H3 was announced as open-weight on 31 July. The weights had not shipped by mid-August.
H3 was presented as a general-purpose omni-modal generation model with video output at up to 2K and native stereo audio, described as open-weight. Two weeks on, the weights were still not available — a gap between announcement and release that is becoming a pattern.
The distance between "open-weight" as an announcement and open weights as a downloadable artefact has become large enough to need tracking separately. H3 is the current example, and it is not the only one.
The model as described is substantial: joint understanding across text, image, video and audio, output at up to 2K with native stereo sound, 4 to 15 second clips. Those capabilities were demonstrated and are not in dispute. What has not happened is the release of weights anyone can run.
This matters because the two claims do different work. A demonstrated capability affects the competitive narrative; released weights affect what people can actually build. Coverage tends to collapse them, and the announcement gets the attention while the release — or its absence — does not.
It is the second instance inside a month, after a Qwen model committed for a specific week that has no repository or date. Neither is evidence of bad faith; staged rollouts are normal. Both are reasons to record the ship date rather than the announcement date.
Hugging Face — What Is MiniMax H3 (Hailuo 3.0)? The Open-Weight Multimodal Video Model, Explained → · The Open Weights — Open-source AI, tracked daily →