According to reports, Alibaba's AI video large model HappyHorse is expected to be released in a week

Alibaba-W (09988.HK) AI video large model HappyHorse has been internally launched on Alibaba Baichain, and it is expected to be officially released to the public in a week, according to domestic media reports. The model ranks first in the ArtificialAnalysis video competition arena leaderboard with a score of 1333 Elo, achieving an 80% win rate against OVI1.1 and a 60.9% win rate against LTX2.3, making it the highest-ranked open-source video generation model globally.

It is reported that the model was developed by a team led by Zhang Di, former Vice President of Kuaishou-W (01024.HK), and has officially announced its open-source status. In the fields of text-to-video (without audio) and image-to-video (without audio), it surpasses Seedance2.0 and Kegling 3.0, slightly leading in text-to-video (with audio), and is on par with Seedance2.0 in image-to-video (with audio).

According to the information, HappyHorse1.0 is currently the world's first open-source video large model that natively supports the joint generation of audio and video, with 15 billion parameters, utilizing a 40-layer unified self-attention Transformer architecture. Generating a 5-second 1080p video on a single H100 takes approximately 38 seconds, and it natively supports lip-syncing in seven languages: English, Mandarin, Cantonese, Japanese, Korean, German, and French, with the lowest word error rate among similar open-source models