MiniMax H3-Image Launches: Native 4K, Open Weights

On November 10, 2026, MiniMax released its first image model, H3-Image: one checkpoint for text-to-image and image editing, native output of about 16.8 megapixels (4096×4096) with no upscaler, distilled to six sampling steps, with up to 16 reference images. The announcement, weights, API and native ComfyUI support all landed the same day — but the story is longer than "launch day": the repository was created a week earlier and sat empty for seven days; the text encoder is reused wholesale from July's H3 video model; and a week before launch, a third party shipped a distilled build whose name differs by a single letter P.

Launch day: four moves in two hours

In UTC, November 10 unfolded like this: at 14:15, MiniMax posted the announcement on X — "16 million pixel. 10-20 seconds. Open weights."; at 14:50, the weights went public on Hugging Face under a license agreement signed the same day; at 16:05, the ComfyUI blog announced native support, with v0.45.0 shipping the same day; at 16:23, the official blog published, titled "Bringing 4K Image Generation into the 10-Second Era." The API release notes are dated the same day. In Beijing time, the blog published at 00:23 on November 11 — so "released November 11" is a timezone artifact; this site uses November 10 as the release date.

Notably, the Hugging Face repo was actually created on November 3 at 06:34 UTC and sat empty for a week before the weights appeared. That's not a leak, just an early placeholder — but the same day, the third party Pruna released P-H3-Image: the H3 video model distilled into an 8-step image model, 8MP, about 21 GB in 4-bit. The names differ by one letter P and the releases are a week apart, yet they are entirely different things; P-H3-Image even got native ComfyUI support first (v0.44.0, November 4).

Where it came from: the H3 video model without the audio branch

H3-Image is not a new model trained from scratch. Its base is H3-Base — MiniMax's video model whose weights landed on July 28 — with the audio branch removed, specialized for still images. The text encoder is reused wholesale: H3's full Qwen3-VL encoder, unchanged, with the DiT reading the hidden states of its 50th layer. That has a practical benefit: users who already run H3 locally don't need to re-download the 66.7 GB text encoder.

Architecturally, H3-Image is a 33.1B dense diffusion transformer: 50 layers, hidden 5,376, 56 attention heads, FFN 14,336, 3D multimodal RoPE. Note that 33.1B counts the DiT only, not the text encoder. The VAE is a CNN encoder plus a 24-layer 2D ViT decoder, 16× downsampling, 24 latent channels; the DiT patchifies the latent with a 2×2 patch, so each token covers 32×32 pixels. For the full spec sheet, see our specs and architecture article.

Three speed wordings, noted here

The launch materials give H3-Image's speed in three wordings that don't fully agree: the blog's headline and text say 4K in "under 10 seconds"; the X announcement says "10–20 seconds"; and the comparison chart in the official blog gives a 4K text-to-image P50 median of 16.0 seconds for Standard and 10.3 seconds for Highspeed. 1K/2K text-to-image on Highspeed genuinely runs under 10 seconds (7.7/7.6s), so "under 10 seconds" only fails at 4K. How the three wordings reconcile, what the chart footnote says, and where "up to 5×" comes from — see our 10-second claim fact-check.

Where it sits in MiniMax's lineup

H3-Image is the third piece of MiniMax's second-half-2026 lineup: the H3 video model on July 28 (weights), the M3.1-Flash language model on October 17 (weights), and the H3-Image image model on November 10. All three share a technical base — H3-Image's text encoder comes from H3, and the PE prompt-expansion model was fine-tuned from Qwen3.8-27B — MiniMax is splitting one multimodal base into video, language and image lines.

There are three entry points: the MiniMax API Platform (platform.minimax.io globally, platform.minimax.cn in China), the Hailuo AI app, and MiniMax Code. Local deployment runs on SGLang, vLLM-Omni, diffusers, LightX2V and ComfyUI. Three days after launch, the main repo had 437,039 downloads on Hugging Face (evening of November 13) — a high number for image-model weights, but downloads are not a quality score.

First-day reaction: busy, but nobody actually measured anything

The coverage from November 10–11 shares one trait: nobody ran an independent test. CellCog did the most complete launch-day fact-check, verifying the X post, API guide, price list and model card item by item; OrcaRouter's judgment was sharpest — ten seconds is architecturally plausible (six unguided steps genuinely sit in a different cost class from a thirty-step guided sampler), but "plausibility is not evidence": the model card states no generation time anywhere, no third-party timing existed as of November 11, there is no AA image-board entry, no GenEval, DPG-Bench or human-preference scores. XenoSpectrum was the first to extract every number from the official chart, while cautioning that whether the blog claim and the chart medians were measured under the same conditions is unknown.

So H3-Image's true state on launch day: specs and prices fully public and checkable, speed and image quality entirely without independent evidence. What to watch next: third-party 4K timing, whether the AA image boards list it, pricing after the 4K half-price offer ends on November 17, and licensing progress in the excluded territories (US, EU, UK, Korea). This site will keep tracking it.