LMArena Rankings — April 2026
HappyHorse 1.0: The Mystery AI Video Model That Dominated LMArena Then Vanished
In early April 2026, HappyHorse 1.0 appeared on the Artificial Analysis Video Arena, instantly topped both text-to-video and image-to-video leaderboards — beating Seedance 2.0, Kling 3.0, and PixVerse V6 — then disappeared within days. Here is the complete analysis.
Architecture Deep-Dive: 40-Layer Single-Stream Transformer
Traditional video generation models use multi-stream architectures where text, video, and audio each have separate encoders interacting via Cross-Attention. HappyHorse 1.0 collapses this into a single pipeline.
Single-Stream Self-Attention Replaces Multi-Stream Complexity
One 40-layer Self-Attention Transformer processes text, video, and audio tokens simultaneously — no Cross-Attention, no modality-specific sub-networks. All modalities are unified into a single token sequence attending to each other in the same attention space.
- Higher parameter efficiency — no redundant parameters for modality isolation
- Shorter inference path — no cross-modal data transfers, more continuous kernels
- Unified training objective — text, visual, and audio share one loss function for end-to-end optimization
- Native audio-video joint generation — sound and visuals are tokens in the same sequence with built-in synchronization
8-Step Denoising + No CFG: Extreme Inference Speed
Only 8 denoising steps with no Classifier-Free Guidance needed to produce Arena-leading quality. This implies Consistency Distillation, Rectified Flow, or Progressive Distillation during training — compressing multi-step sampling into direct prediction. Combined with the released distilled and super-resolution models, the full stack targets both edge-friendly and high-throughput server deployment.
Estimated Parameter Count
Weights are not yet public, but given the 40-layer single-stream architecture, 6-language support, and Arena performance, the model likely falls in the 10B–30B parameter range — comparable to Wan 2.x, Seedance 1.x, and Hunyuan Video.
