MiniMax a lancé H3 le 31 juillet. La combinaison est rare: un prix d'API commercial et un plan d'ouverture des poids du modèle.
H3 accepte texte, images, vidéo et audio dans un même contexte. Il génère jusqu'à 2K et 15 secondes, avec audio stéréo natif.
L'ouverture reste une promesse
“We plan to open up the model weights in the coming days, subject to applicable laws and regulations.”
Developers still need the license, commercial-use terms, hardware guidance and a technical report before treating H3 as a deployable open model.
Le prix rend l'attaque concrète
Shanghai Securities News, The Beijing News and 21st Century Business Herald reported H3 at RMB 0.8 per second for 2K generation, about one-third of comparable flagship video models. A 15-second 2K clip is roughly RMB 12 before platform packaging.
Le pari technique est le contexte unifié
MiniMax describes H3 through Contextual Omni Representation, H3-VAE, H3-Omni Transformer and In-Context Regeneration. The model handles video, audio, images and text as one creative context.
Most source material requires about 100,000 tokens of inference before being distilled to roughly 4,000 tokens on average. H3-VAE provides a 4x gain in effective sequence length, and workload separation lifted training throughput by nearly 30%.
The next checks are whether weights appear, whether the license works for companies, and whether the community can reproduce usable text, motion and audio control on non-official hardware.
Sources: blog de recherche officiel de MiniMax, Shanghai Securities News, The Beijing News, CocoLoop, 21st Century Business Herald; site MiniMax pour vérifier 2K/15 secondes, stéréo native, engagement d'ouverture des poids, prix de 0,8 yuan par seconde, comparaison 768p et gain de débit d'entraînement proche de 30%.