China Steals the Leaderboard, OpenAI Upgrades Images, and the Generative Media Race Enters a New Phase

News

Alibaba's Stealth Model HappyHorse-1.0 Debuts at #1 on Every Text-to-Video Leaderboard

HappyHorse-1.0 appeared anonymously on Artificial Analysis around April 7, immediately claiming the top spot in both text-to-video and image-to-video blind rankings — then Alibaba revealed it was behind the project four days later. The model is still in closed beta with no public API, but the fact that it unseated ByteDance's Seedance 2.0 and Kling 3.0 in blind tests signals that China's intra-lab video war is producing serious quality jumps, and open-source weights are reportedly coming soon.

ByteDance Rolls Seedance 2.0 Into CapCut Globally After Hollywood Copyright Standoff

After a forced pause triggered by MPA cease-and-desist letters and a Japanese government probe, ByteDance relaunched Seedance 2.0 inside CapCut with face-generation blocks, invisible watermarks, and C2PA credentials — but the US market is still excluded. Embedding the model directly into CapCut gives ByteDance a massive distribution edge over standalone tools, though the 15-second clip ceiling is a hard limit for anyone outside short-form social content.

Analysis

HappyHorse's Creator Zhang Di Was the Architect of Kling AI — and He Just Beat His Own Work

The story behind HappyHorse is as interesting as the benchmark numbers: Zhang Di, the former VP who built Kuaishou's Kling AI, rejoined Alibaba in November and led a months-long sprint to ship a top-ranked video model under a pseudonym. It's a pointed reminder that individual talent and internal lab structure matter as much as compute in the current video generation race.

What HappyHorse's Leaderboard Climb Actually Means — and What's Still Unverified

This breakdown from WaveSpeedAI cuts through the hype: HappyHorse holds a 60-point Elo lead over Seedance 2.0 in silent T2V, but its vote sample count is still small relative to established models, weights are unavailable, and no independent technical audit has confirmed the claimed 15B parameter count or 38-second 1080p generation speed. Worth bookmarking as a template for how to evaluate any new stealth video model drop in 2026.