Alibaba HappyHorse-1.0: The Mystery Model That Topped Every AI Video Leaderboard

The Reveal

Artificial Analysis runs a Video Arena where users submit a prompt, two anonymous models generate outputs, and users pick the one they prefer. Votes feed into an Elo rating system (the same math used in chess rankings). Models cannot game the system because evaluators never know which model produced which output.

HappyHorse-1.0 entered this arena on April 7 under a blank profile. No logo, no company attribution. By April 10 it held the top position in two of the four ranking categories, and Alibaba confirmed ownership through a newly created X account and a statement to CNBC.

The Numbers

HappyHorse's Elo scores tell the story clearly:

For context, the previous #1 in text-to-video was ByteDance's Seedance 2.0 at Elo 1273. HappyHorse beat it by 60 points, a gap that typically takes months to close. In image-to-video, HappyHorse scored 1392 versus Seedance 2.0's 1355.

The audio-inclusive categories tell a different story. HappyHorse sits at #2 behind Seedance 2.0 (Elo 1219 vs. 1205), suggesting the audio pipeline still needs work relative to video quality.

Architecture: One Transformer, Everything at Once

Most AI video systems chain separate components: a text encoder, a video generator, and an audio model bolted on afterward. HappyHorse takes a different approach.

The model uses a 40-layer single-stream Self-Attention Transformer with no Cross-Attention modules. Text, video, and audio tokens all flow through the same transformer stack simultaneously. This unified design eliminates redundant parameters and reduces inference complexity.

The inference pipeline is unusually lean: only 8 denoising steps with no Classifier-Free Guidance (CFG). For comparison, many competing models use 25-50 steps with CFG enabled. This suggests aggressive distillation during training, trading raw capacity for speed.

The Team Behind It

HappyHorse comes from Future Life Lab under Alibaba's Taotian Group (the e-commerce arm). The lead is Zhang Di, former VP of Kuaishou and technical lead of Kling AI.

That detail matters. Zhang Di built the engineering culture behind Kling, one of the strongest AI video models of 2025. He knows the infrastructure challenges, the training pipelines, and the evaluation gaps. Bringing that experience to Alibaba's compute resources is a different equation than running a standalone startup.

Why This Matters for the Market

The AI video market in April 2026 is unusually volatile:

With Sora winding down and Seedance facing legal headaches, HappyHorse arrives at a moment when the market is looking for a new front-runner. Runway Gen-4.5 still holds strong in production workflows, and Google Veo 3.1 dominates the enterprise integration space. But for raw generation quality measured by blind human preference, HappyHorse now sits on top.

Open Source: Coming Soon (Maybe)

The GitHub repository and Hugging Face model hub both show "Coming Soon" as of April 11. The team has promised open-source release of the full 15-billion parameter weights with a commercial license. If that happens, HappyHorse would be the largest open-source video model available, significantly bigger than LTX-Video's 2B parameters.

But "coming soon" is not "released." Until weights are downloadable and independently verified, the benchmark results carry an asterisk. The AI video community has learned to wait for reproducible results before declaring winners.

What to Watch

Three things will determine whether HappyHorse maintains its lead:

• ✓Open-source weight release and independent reproduction of benchmark results
• ✓Audio quality improvements to match or beat Seedance 2.0 in the audio-inclusive categories
• ✓API availability and pricing, since benchmark dominance means nothing if creators cannot access the model

The AI video generation space moves fast. In January, Runway Gen-4.5 looked untouchable. In February, Seedance 2.0 took the crown. Now HappyHorse holds it. The question is not whether something will eventually beat it, but when and from where.

For creators and developers building video pipelines today, the takeaway is practical: no single model stays on top for long. The best strategy is to build workflows that can swap models as the leaderboard shifts, rather than betting everything on one name.

Sources

• Kaiyuan Securities Research Institute: HappyHorse-1.0 reached 1333 Elo and topped the Artificial Analysis video leaderboard on April… (Kaiyuan Securities Research Institute)