Where the open versus closed split goes

August 2026 gave the open-versus-closed question a third answer that neither side had prepared for. MiniMax released H3, a frontier-class video model with native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward., as open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models., and within a month the ecosystem did exactly what the pattern predicts: a research lab distilled it to four steps, an inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. provider post-trained it into a product that outranked the original on a public arenaArenaA public leaderboard that ranks models from large numbers of blind pairwise human preference votes, which tends to reward average appeal over fitness for a specific job., another lab shipped a ControlNetControlNetA conditioning technique that lets you provide structural guidance (depth, edges, pose) to a diffusion model to control spatial layout. for it, and a training toolkit added LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. support. That is the Stable Diffusion story running again, at video scale, in weeks rather than months.

The difference is in the license. The MiniMax H3 Community License limits its grant to an Applicable Territory defined as worldwide excluding the European Union, the United Kingdom, the Republic of Korea, and the United States of America, and it extends that exclusion to outputs, not just to the weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable.. Parties in the excluded territories are invited to negotiate separately. The derivatives inherit the terms, so the four-step distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.). and the LoRA adapters carry the same border the base model does. Compare MiniMax's own music model, published the same month under an attribution-style grant with no territorial restriction at all, and it becomes clear this is a deliberate per-model decision rather than a house policy.

Most people's intuition runs the other way. Video is the more expensive medium to train and to serve, so you would expect open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models. to trail hardest there and to have closed the distance on still images, which are cheaper by every measure. The likely explanation is not compute. It is who is choosing to publish. The labs putting frontier video weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. into the world, MiniMax, Lightricks, Sand.ai, and Alibaba through the Wan 2.x line, are challengers using open release to buy distribution in a market where they are not the incumbent. The labs at the top of the image board, OpenAI, Microsoft, Google, and Meta, all keep their weights closed, and none of them has a reason to do otherwise. Reve belonged on that list until September 2026, when it shut its generation service down rather than open anything.