On 10 November 2026, MiniMax published MiniMax-H3-Image on its Chinese official website. The official blog post is titled "MiniMax H3-Image 开源：让 4K 图像生成进入 10 秒时代." An English official blog post at the same path was not up.
The visible date on the Chinese official blog is 10 November 2026. Separately, the blog's JSON-LD datePublished and dateModified are 2026-11-10T15:00:58.000Z, which is 23:00:58 Asia/Shanghai. MiniMax's official account post has a created_at of 2026-11-10T15:40:15.000Z, which is 23:40:15 Asia/Shanghai. These are two official clocks, recorded separately.
MiniMax-H3-Image is an image generation model derived from MiniMax H3. The Chinese official blog says it has text-to-image, image-reference, and image-editing capabilities, and that the weights are being opened. The official ID for the high-speed version is MiniMax-H3-Image-Highspeed.
On availability, the Chinese official blog says that besides opening the model weights, MiniMax is also providing H3-Image image generation through the MiniMax open platform, MiniMax Code, and MiniMax Hailuo. The official account puts it this way: "Available now via our API, MiniMax Code, and Hailuo AI." That is the official scope as written.
Weights are on Hugging Face at MiniMaxAI/MiniMax-H3-Image. The license title is MiniMax-H3-Image COMMUNITY LICENSE AGREEMENT, dated November 10, 2026 — a date only, with no time of day. The applicable territory excludes the European Union, the United Kingdom, the Republic of Korea, and the United States. If annualized revenue exceeds 1 million USD, or if a Hosted Service is offered to third parties, separate written authorization is required. This is a community license for open weights, not unrestricted open source, even though the Chinese blog title uses "开源" and the Hugging Face README calls it an open-source model.
The Chinese official blog says the model can generate a 4096 × 4096, about 16 MP (4K) image in 10 seconds, and that it supports up to 16 reference images. It also says a 4K image can be generated in 10s at up to five times the speed of mainstream levels, that standard mode is 1.5-2 times as fast as mainstream tools on the market, and that Highspeed mode can reach 3–4 times. It further says the price is 50%–70% of comparable mainstream tools, and that the model offers the strongest instruction following among open-source models. Those statements come from the official blog; they are not independent measurements.
The Hugging Face model card says generation and editing are unified in one checkpoint, that the DiT has 33.1B parameters, that the model works at about 16 megapixels, and that it supports up to 16 reference images. The model card does not say "10 seconds." The "unified in one checkpoint" wording comes from the Hugging Face model card. The Chinese official blog describes the same three capabilities.
The API documentation, the promotional wording, and the discount threshold use three different pixel figures. The API documentation says the output area is up to 20,000,000 pixels. The promotional wording is about 16 MP (4K), or 4096 × 4096. The discount threshold is 16,600,000 pixels.
Official latency comparison
The Chinese official blog includes an official latency comparison chart, titled "不同清晰度下的模型耗时对比." It is a latency comparison, not a quality benchmark, and not a third-party public leaderboard.
The official chart reports P50 (median latency) in seconds; lower values mean shorter latency. It is organized into text-to-image, image editing with 1 input image, and image editing with 3 input images, each shown at 1K (1024×1024), 2K (2048×2048), and 4K (16MP, 4096×4096). The comparison includes H3-Image Standard, H3-Image Highspeed, and several other models; the 4K groups do not draw every competitor bar. The official caption on the chart reads: "数据来源：用户提供的对比图及二次评测表。H3 采用二次评测 P50；其他模型沿用原图。各图纵轴统一为 0-70 秒。" That is: the data come from user-provided comparison charts and a secondary evaluation table; H3 uses the secondary-evaluation P50; other models follow the original chart; and every panel uses a 0–70 second vertical axis. On the 4K text-to-image panel, the chart's P50 values are 16.0 seconds for the standard version and 10.3 seconds for Highspeed. Those chart figures are separate from the official blog's "10 seconds" and "10s" wording.
Token pricing and the limited-time discount
Pricing is token-based, with separate rates for text input, image input, and image output, per million tokens.
On the international pricing page, MiniMax-H3-Image (Standard) is $2.30 / $1.50 / $3.80 per million tokens for input text, input image, and output image. MiniMax-H3-Image-Highspeed is $2.30 / $3.00 / $7.60. On the domestic pricing page, the standard version is 15 / 10 / 25 yuan per million tokens, and the high-speed version is 15 / 20 / 50 yuan.
The international pricing page states that from now until November 17, 2026, at 16:00 UTC (exclusive), the total request cost for MiniMax-H3-Image and MiniMax-H3-Image-Highspeed is 50% off when the final delivered image has at least 16,600,000 pixels (width × height). The domestic pricing page gives the same window as running through 18 November 2026, 00:00 Beijing time, exclusive of that end instant, at 50% off the total request cost when width × height is at least 16.60 million pixels. The cutoff is 2026-11-17T16:00:00Z (exclusive), which is 18 November 2026, 00:00 Beijing time, exclusive. The official account's "launch week" and the Chinese official blog's wording about the first week after release do not replace that pricing-page cutoff.
Example per-image costs on the pricing pages are for reference only; they are not fixed per-image prices. The international examples are rounded to 5 decimal places. The domestic examples are rounded to 4 decimal places, and the domestic page also notes that some 1K/2K example amounts have been adjusted for presentation. The two tables use different currencies and different decimal-place rules.
The Chinese official blog also includes a "单张生成费用参考" table. That table is a cost reference in yuan, not a benchmark and not a fixed per-image price. Some 4K / 16MP rows are marked with a limited-time 50% discount, shown as a struck-through original price and a promotional price.
The official blog cover, Hugging Face showcase images, and the official account's demo video are official examples or visuals, not evaluation charts.
Hugging Face also hosts an optional companion repository, MiniMaxAI/MiniMax-H3-Image-PE, described as a prompt-expansion companion. It is not a second image-generation weight release.