A split result, clip by clip
The clearest gap was camera work. Asked to push in and tilt up toward the trophy's glowing tip, Wan did both and ended on the tip. H3 barely moved; only the flare at the tip grew. Asked for lines of light branching inside the crystal, Wan drew them from base to tip. H3 showed them faintly, in the last second or two.
H3 won on the material itself. In the wind prompt, its cloth billowed in large waves while Wan's moved less. In the spreading-light prompt, H3 sent radial streaks across the cloth and kept the trophy large, where Wan drew concentric rings, pulled far back, and made the text unreadable. Across the ten clips, the source rates H3 best at keeping logo and text intact, readable throughout almost every clip.
Both models generated sound for every clip, including prompts that said nothing about sound. The source's practical advice is to plan on replacing that audio. Both also followed the three-shot structure of the 10-second prompt, but treated the first frame differently: Wan started from the photo's wide composition, while H3 showed the photo for one frame and then cut to the close-up the script asked for.
Read the setup before the scoreboard
Three conditions tilt this test, and the source states all three itself.
First, the prompts were written the Alibaba way. They follow Alibaba Cloud's official "motion + camera movement" formula, and the source says plainly that they are not tuned for H3. A test written in one model's prompt dialect measures dialect fit as well as model quality.
Second, the two models did not run in the same place. Wan ran on a cloud API at 720p, three jobs in parallel. H3 ran locally, one job at a time, at a different resolution, on a single 20 GB card, with one seed (20261113). The speed gap is real for this setup. It is not a controlled measure of the two models on equal hardware.
Third, the publisher sells access to one of the two. AICU sells API access to Wan, and the article closes with a call to run Wan on the AICU API, with pricing in API points (1,400 AP per second at 720p; 77,000 AP for all ten clips, as of publication). That does not make the frame strips false; they are published for anyone to inspect. It does mean the test itself, a single trophy photo, a single subject, no people, no dialogue, is worth reading as one interested party's choice of ground.
What to take from it
For a product or logo shot where the text has to survive, this is useful evidence for H3 in this kind of shot, with a cost attached: 481–501 seconds a short clip on the test's 20 GB card, and a community license that restricts display of outputs in several major markets, including the United States. For a teaser built on a specific camera move, Wan's obedience and its measured times on the API are the point, and AICU's own workaround is worth noting: pass the original image as the last frame as well, so the clip ends on the untouched logo image. That controls the ending frame; it does not keep text legible through the moves before it.
What this test does not settle is ranking. One photo, one prompt dialect, one seed and two different runtimes cannot rank two models. In AICU's trophy test, Wan followed the requested camera moves more closely, while H3 better preserved the logo and text. Neither setup establishes which model is better in general, and neither says anything about people, dialogue or multi-character scenes, which the test does not contain.
Sources
- [1] AICU, "Same 10 prompts, two models — Wan 3.0 on the API vs MiniMax H3 on a local GPU," November 21, 2026Read source
- [2] Comfy-Org/MiniMax-H3 model files on Hugging FaceRead source
- [3] MiniMax H3 Community License Agreement, via the MiniMax-H3 repository on Hugging FaceRead source