news

MiniMax opens H3-Image weights with 16 megapixels in six steps, and no measured latency

The 33.1 billion parameter checkpoint renders up to 6272 by 2688 without an upscaler, while the "4K in ten seconds" line arrives with no named hardware and no quality score.

By Priya Shenoy · 15 November 2026

MiniMax published the weights for H3-Image on November 10, and the specification sheet is considerably more precise than the speed claim stapled to it. The model card describes a 33.1 billion parameter diffusion transformer distilled to six sampling steps with no classifier-free guidance, producing roughly 16 megapixels directly. The company's launch line, 4K image generation in ten seconds, arrives with no named GPU, no batch size, and no measured timing to support it.

One checkpoint for generation and 16-reference editing

H3-Image handles text-to-image and reference-based editing from the same file, accepting as many as 16 reference images in a single edit. Output sits at about 16 megapixels across supported aspect ratios, from 4096 by 4096 at 1:1 up to 6272 by 2688 at 21:9, generated without a separate upscaling pass. The six-step schedule means the transformer runs six neural function evaluations per request, and because guidance is fixed at 1 there is no classifier-free guidance pass and therefore no negative prompt.

The architecture is reused rather than newly designed. H3-Image is the MiniMax H3 video model with the audio layers stripped out, retaining the Qwen3-VL text encoder from that release, which is why the download list pairs a 66.28 GB transformer with a 66.7 GB encoder pulled from the H3 repository. The video sibling is already on our price radar across nine platforms.

Ten seconds is a claim, not a result

What the release does not carry is a latency table, the hardware any timing was produced on, or a single quality score placed against another image model. MiniMax describes the 16 megapixel render as roughly five times faster than the norm for that resolution, which is a comparison without a stated baseline model, resolution, or sampler configuration on the other side of it. Six steps at 16 megapixels is genuinely cheap compared to a guided 30-step run at the same pixel count, so the direction of the claim is plausible. The magnitude is still unverified.

That pattern is becoming routine for open-weight image releases. Black Forest Labs shipped FLUX 3 Image Klein earlier this month with every first-place claim scored in-house, and H3-Image goes one step further by publishing no comparative scores at all.

Open weights with a territory attached

The weights are hosted on Hugging Face under a MiniMax-H3-Image Community License rather than a standard permissive license, and secondary write-ups do not agree on how far its territorial exclusions reach. One account describes the license as excluding United States and European Union use of the weights; another counts four excluded major markets. Anyone planning commercial deployment should work from the license text rather than any of the summaries circulating around it.

Day-zero tooling followed quickly, with ComfyUI support landing in v0.45.0 and recommended settings of euler, simple scheduler, six steps, CFG 1 and denoise 1. Hosted access runs through the MiniMax API and Hailuo AI alongside third-party hosts.

Our own scoring of the H3 video model, which shares the backbone, is below for anyone trying to infer what the image branch inherits.

The useful question over the next few weeks is whether anyone outside MiniMax publishes a wall-clock number at 4096 by 4096 on named hardware. Until then, 33.1 billion parameters, six steps and 16 reference images are facts, and ten seconds is marketing copy.

Sources: MiniMax (Hugging Face model card), Pandaily.

Priya Shenoy: Tracks what AI video actually costs across the platforms that resell access to the same handful of models. Treats a pricing page as a claim, not a fact, until someone checks it.