On November 10, 2026, MiniMax released H3-Image, an open-weight image generation model built on H3, leading with a single number: a 4K image at 4096×4096 in under 10 seconds. The model supports text-to-image generation, reference-based generation, and image editing, and its weights are open, so developers and creators can download and use it directly.
Speed is the hardest number in this release. MiniMax says H3-Image generates 4K images at up to 5 times the speed of mainstream alternatives; Standard mode runs 1.5 to 2 times faster, and High-speed mode reaches 3 to 4 times faster. The speed comes without cutting quality: details look closer to real photographs instead of the overly smooth "AI look", prompt adherence ranks at the top among open-weight models, and the model can understand up to 16 reference images at once. There is no single trick behind it; according to MiniMax, the gains came from improving the entire post-training pipeline, from data and reinforcement learning to sampling-step reduction, positional encoding, and deployment optimization.
Pricing came down with the speed. Official prices are 50% to 70% of comparable mainstream offerings, with a further 50% off image generation requests at 4K resolution and above during launch week. There are three ways in: the MiniMax API Platform, MiniMax Code, and Hailuo AI, and the weights are on Hugging Face under the organization MiniMaxAI, project MiniMax-H3-Image. For teams that produce images at volume, this means 4K output can run as part of the everyday workflow for the first time, instead of being saved for one final render.
Open weights are the biggest difference between H3-Image and closed image models: the model can be customized and wired into your own tools and workflows instead of waiting on official updates. The trade-off is that some of the deployment and tuning work falls to you, since the official release provides the weights and API access. MiniMax took the same route when it opened the weights for M3.1 Flash: MiniMax open-sources M3.1 Flash weights: 428B parameters with only 23B active, adapted to domestic chips on day one.
