MiniMax open-sources H3-Image, a 4K image model with 16-reference input and cut-rate API pricing

MiniMax has open-sourced H3-Image, a new image model that follows its open-source video flagship H3, according to a hands-on review published by Chinese tech outlet QbitAI on November 12. The piece runs through a long series of generated samples and argues that the company has become one of the most capable players in multimodal generation.

The headline specification is resolution: native 4K at 16MP, or 4096 by 4096 pixels. In the reviewer's hands a full 4K image took roughly 10 to 20 seconds on the web interface, and the model accepts up to 16 reference images, a feature aimed at character consistency and continuity across complex storyboards.

The more interesting capability is structural understanding. Given a photo of a product, H3-Image can derive a standard orthographic three-view drawing with front, side and top views. Asked for an exploded mechanical diagram, it layers the parts along assembly logic and labels the components with callout lines — work that depends on three-dimensional geometry and spatial reasoning as much as on rendering.

Text layout is the other dimension the review stresses. The model produced a calculus exam paper, complete with limit notation and complex integrals, in about ten seconds without garbled characters, at a density close to a real printed page. For commercial work, whether the information survives intact matters more than whether the picture looks pretty.

On realism, the review argues that H3-Image largely clears the plastic look of earlier image models. Pores, freckles, fine facial hair, water stains and dust on a mirror, and the flare of a flash hitting a lens all come through. The author credits these deliberately imperfect details for making the output feel closer to human-made work.

Openness and price are the second half of the signal. The model is open source, so developers can customise and deploy it, and API calls are priced at 50% to 70% of mainstream competitors, with a further 50% off 16MP requests in the first week. The headline's figure of two mao, roughly 0.2 yuan, per image comes from that pricing.

The reviewer draws a broader lesson: rather than pushing one general model to do everything, wiring specialised models into a harness and letting an agent call them as tools may deliver more total utility. On that reading, a dedicated image model handles asset generation while a general model orchestrates the workflow.

If those three things hold — less AI sheen, generalisable taste and reliable logic — multimodal generation may be having its own coding moment. Industrial packaging design, commercial illustration and multilingual cross-border marketing assets are all workflows whose timelines could compress sharply. What to watch next is how the open weights perform in community deployments, and whether the low pricing holds.