MiniMax H3-Image 10-Second Claim vs Its Own Chart

The loudest slogan at H3-Image's launch was "4K images in under 10 seconds." But the comparison chart in the same official blog reads: 4K text-to-image P50 median, 16.0 seconds for Standard, 10.3 seconds for Highspeed — even the fast tier exceeds 10 seconds. The claim and the chart, one source, two numbers. This article lays out all three speed wordings, the chart footnote and the origin of "up to 5×," and checks which one holds.

Three wordings, three numbers

The X post's "10–20 seconds" matches the chart: 16.0 and 10.3 both fall in the 10–20 second range. The blog's "under 10 seconds" directly conflicts with the chart: both tiers' medians exceed 10 seconds. XenoSpectrum was the first to extract every number from the chart, and flagged a key point — whether the blog claim and the chart medians were measured under the same conditions is unstated by MiniMax, so the difference alone doesn't prove either side "wrong." But as a reader, you can be sure of one thing: by the official chart's own measure, the 4K text-to-image median is not under 10 seconds.

Every number in the chart

The pattern is clear: higher resolution is slower, more reference images is slower; Highspeed is about a third faster than Standard. Notably, 1K/2K text-to-image on Highspeed genuinely runs under 10 seconds (7.7/7.6s) — "under 10 seconds" holds at 1K/2K and only fails at 4K. The blog headline's combination of "4K" and "10 seconds" is precisely the pairing that doesn't hold together.

The chart footnote: not a same-environment speed ranking

The official chart's footnote says H3-Image figures use "second-round P50" while other models use original data. That means the chart is not a speed ranking measured in one environment — H3-Image's numbers come from a second round of measurement (typically a warm start, after cache warm-up), while rivals' numbers are their own original reports. The 4K column also has only Nano Banana 2.1 as a rival (GPT-image 2.5/2.6 and Seedream 5 Pro have no 4K values). So the chart can show "H3-Image is much faster than Nano Banana 2.1," but it can't show it's fastest against all rivals.

Three more unpublished conditions: hardware (which GPU, how many), prompts (how long, how complex), and timing scope (whether pre- and post-processing are included). P50 is a median, not a guarantee — half of requests are slower than the median. OrcaRouter notes the model card states no generation time anywhere and no third-party timing existed as of November 11; the entire evidence chain for the speed claim is currently MiniMax's own materials.

Where "up to 5×" comes from

The blog claims H3-Image is "up to 5×" faster than competitors. Working back from the chart: Nano Banana 2.1 medium 4K text-to-image at 50.0 seconds divided by H3-Image Highspeed at 10.3 seconds is about 4.9× — that's the origin of "5×" (this site's calculation; MiniMax hasn't stated the method). The multiple only holds for 4K text-to-image against Nano Banana 2.1 medium; change the task or the rival and the multiple changes. Reading "up to 5×" as "5× faster across the board" is a misreading.

Conclusion: when "under 10 seconds" holds

By the official chart's measure: 1K/2K text-to-image on Highspeed holds (7.7/7.6s); 1K/2K edits on Highspeed sit around 10 seconds (10.0/10.0/12.0/11.2s — 3 references exceed it); 4K never holds (fastest is 10.3s). The blog's "4K in under 10 seconds" doesn't match the official chart's medians; the X post's "10–20 seconds" does. The final judgment needs third-party timing — as of November 13, there is none.