MiniMax released its H3-Image generation model late on 10 November and published the weights at the same time, putting a frontier-class image model into the open.
The headline capability is speed. The fast tier of H3-Image can produce a 4096 by 4096 pixel image, about 16 megapixels, in roughly ten seconds, which the company describes as about one fifth the time mainstream models require for comparable output. The standard tier runs in the 10 to 20 second range depending on resolution and complexity.
What was actually tested
Zhidx connected to H3-Image through MiniMax’s Hailuo interface and its Code harness and ran a set of structured evaluations across visual fidelity, complex task understanding and multi-reference fusion.
On photorealism, the model produced portraits with natural skin texture and minor blemishes rather than the over-smoothed, over-sharpened look common to earlier generation systems. In a nature scene, the motion of gulls over breaking waves and the layering of the water held together without obvious structural errors. In architectural images, interior ceiling lighting and exterior reflections coexisted without contradicting each other.
Style transfer was tested by converting a single cat photograph into felt stop-motion animation and into surrealist oil painting. For the felt version, the prompt asked for interwoven fibre texture rather than a simple blur, and the model applied the treatment across every element including the leaves. For the surrealist version, a prompt mentioning Dalí alone was enough for the model to add melting clocks and restructure the scene while preserving the subject.
The professional tasks are the interesting part
Image models usually fail first on text and on spatial consistency. H3-Image handled both better than expected.
A poster reading “come drink coffee” rendered all characters correctly and followed an instruction to keep the design minimal. A harder test asked for a poster containing a headline, six numbered steps, roughly 100 Chinese characters and a set of numerals with units including 15 grams, 92 degrees Celsius, 30 seconds, 2 minutes 30 seconds and 1:15. The title, the six steps and every number and unit came out correct, with no misalignment between icons and their labels.
The 16-reference test uploaded 16 separate assets in one pass, including character references, clothing, hair colour and pose references, material swatches, a colour scheme, a typeface sample, a logo position diagram and a layout wireframe, and asked for a band tour poster. All 16 constraints were checked individually and all were met.
With MiniMax Code acting as the orchestrator, the model produced a mechanical exploded view of a hairdryer, separating the shell from internal components along the assembly direction. It also produced a three-view industrial design sheet from a brief specifying dimensions, shell material, display size, port positions and internal module layout. The only visible error was a slight proportional deviation in the display width, which the harness fixed by having the model write a pixel measurement script to locate and correct it.
A logic puzzle was rendered not as question and answer but as a structured infographic with a conditions block, a stepwise elimination section and a final answer, with a clear information hierarchy.
Why open weights change the arithmetic
For design and manufacturing teams in China, the commercially relevant capability is not artistic quality but reliable text rendering and view consistency. A model that can hold a three-view drawing together and set 100 Chinese characters correctly is usable in a production workflow. A model that produces a beautiful image with garbled labels is not.
Releasing the weights removes two constraints at once. It removes per-image pricing, and it removes the queue. Teams with their own compute can fine-tune on proprietary product imagery and run inference inside their own perimeter, which matters for anyone handling unreleased industrial designs.
The remaining gap for closed models is therefore not capability but distribution, latency at the high end and the final increment of aesthetic judgement.
Editor’s note: This is an adapted translation of the original Zhidx report. It has been trimmed and restructured for readability for an international business audience. The full original (in Chinese) is at https://www.zhidx.com/p/609469.html.