Technology

Add as a preferred source on Google

MemeBench Gives MiniMax H3 Top Score in Narrow Video Preference Test

The leaderboard shows overlapping score ranges for H3 and ByteDance’s Seedance models, and the test covers only meme-style clips.

Updated 2 min read
A viewer compares meme-style video clips on a laptop in a Shanghai media studio. / EconoTimes.ai
A viewer compares meme-style video clips on a laptop in a Shanghai media studio. / EconoTimes.ai

MemeBench’s highest score for MiniMax H3 comes from audience choices on meme-style clips, and overlapping confidence ranges leave the leaderboard unable to establish a decisive lead over ByteDance’s models.

In a leaderboard check dated Oct. 29, 2026, Korea Standard Time, MemeBench listed H3 at 1,423 points, with an approximate 95% confidence interval of 1,344 to 1,502. Seedance 2.0 scored 1,390, with a range of 1,311 to 1,469; Seedance 2.5 scored 1,360, with a range of 1,281 to 1,439. The ranges overlap for all three models.

H3’s tally reflected 10,453 eligible comparisons, compared with 8,685 for Seedance 2.0 and 5,231 for Seedance 2.5. MemeBench’s methodology says visitors can submit multiple ratings, so those totals do not represent unique voters.

The creators’ Aug. 7 post says the first evaluation used 50 prompts based on popular memes from Bilibili and Douyin. Viewers chose between clips generated from matching images and prompts without being told which model made them. MemeBench describes its scores as relative preferences for its current meme scenarios and cautions that its self-selected audience and narrow prompts do not represent every use of the models.

The companies describe different technical capabilities. ByteDance announced Seedance 2.5 on July 31, 2026, saying it can generate videos up to 30 seconds in one pass and highlighting professional uses such as film and advertising. MiniMax announced H3 on July 31, 2026, describing it as a multimodal system able to produce 15-second clips at up to 2K resolution with stereo audio. The leaderboard does not test every feature in those company descriptions.

In an Oct. 29 ChinaTalk interview, Xianyu, the benchmark’s creator, argued that model quality depends on what people are asked to evaluate. He said some AI application developers build products around fundraising rather than clear user needs, and that venture-capital themes partly shape interest in “world models.” In ChinaTalk’s lightly abridged English translation, Xianyu said, “AI isn’t that important, and you don’t have to believe in it.”

Related articles

View all