HappyHorse 1.0

Debuted April 7, 2026 (appearing anonymously on the Artificial Analysis Video Arena on April 8), HappyHorse 1.0 is Alibaba's state-of-the-art AI video model built by the Future Life Lab at Taotian Group, led by Zhang Di — the former Kuaishou VP behind Kling AI. It instantly claimed #1 in both text-to-video (1333 Elo) and image-to-video (1392 Elo), dethroning ByteDance's Seedance 2.0. It is a ~15B-parameter single-stream 40-layer Transformer (no cross-attention, Transfusion-style) that jointly generates video and synchronized audio — dialogue, ambient sound, lip-sync — in a single forward pass, with native 1080p output and ~10-second generation times. The open-source release includes the base model, a distilled faster variant, a super-resolution module, and full inference code.

Limitations: clips are short (5–10 seconds per generation); 4K requires the super-resolution module rather than native generation. As a 15B model it needs serious GPU memory for self-hosting despite the open weights. Third-party API access (e.g. via MuAPI wrappers) adds cost and dependency versus running the weights yourself.

Artificial Analysis: 1333 Elo T2V, 1392 Elo I2V (#1 at launch, April 2026)