HappyHorse-1.0 Deep Dive: Alibaba's 'Happy Horse' Shakes Up AI Video Generation

Forum topic · 小凯 · 2026-04-10

Summary

HappyHorse-1.0, an anonymous video generation model from Alibaba's Taotian Group Future Life Lab, topped the Artificial Analysis video arena leaderboard on April 7, 2026 with an Elo score of 1333, roughly 60 points ahead of ByteDance's Seedance 2.0, ranking first in all four tracks (text-to-video and image-to-video, with and without audio). Built on a 15B-parameter, 40-layer single-stream Transformer without cross-attention, it jointly generates video and audio (lip-synced speech, ambient and Foley sounds) in one pass, supports lip sync in seven languages, and outputs native 1080p. DMD-2 distillation cuts denoising to 8 steps without CFG, generating 5-second 1080p video in about 38 seconds on a single H100. The team is led by Zhang Di, former Kuaishou VP known as the 'father of Kling.' Alibaba claims full open-source release (weights, distilled model, upscaler, inference code, commercial license), though repos still show 'Coming Soon.' Fake websites have appeared, so users should only trust official channels.

Key points

An anonymous model named HappyHorse-1.0 quietly topped the Artificial Analysis video arena leaderboard on April 7, 2026 — and dominated. It scored Elo 1333, nearly 60 points ahead of second-place ByteDance Seedance 2.0, and ranked #1 in all four tracks: text-to-video and image-to-video, with and without audio. As of April 9, it held 1383 in pure video quality (T2V, no audio), a 110-point lead over Seedance. The team behind it is Alibaba Taotian Group's "Future Life Lab," led by Zhang Di, former Kuaishou VP and the creator of Kling.

Technical architecture

The core theme is unification — one model, one pass:

Single-stream Transformer (40 layers): a 4-layer input stage embeds text, image, video, and audio tokens; 32 shared layers process all modalities in one sequence via self-attention; a 4-layer output stage decodes video frames and audio waveforms. No cross-attention — audio-visual sync is generated jointly, not aligned in post.
DMD-2 distillation (Distribution Matching Distillation): denoising compressed from 50+ steps to 8 steps, without CFG (classifier-free guidance). Combined with an internal MagiCompiler runtime, a single H100 generates 5s of 1080p in ~38 seconds (vs. ~3–5 min for Stable Video Diffusion).
Per-Head Sigmoid Gating: a learnable scalar gate per attention head suppresses destructive cross-modal gradients, stabilizing joint training.
Timestep-free denoising: the model infers noise level from the latent itself, removing explicit timestep embeddings and enabling extreme distillation.

Capabilities

Native joint audio-video generation: lip-synced speech, ambient effects, and Foley sounds generated together, semantically aligned with the visuals (footstep rhythm matches walking, rain audio matches on-screen rain).
7-language lip sync: English, Mandarin, Cantonese, Japanese, Korean, German, French — reportedly the lowest word error rate among comparable open models.
Native 1080p output: 5–8 second clips, aspect ratios 16:9 / 9:16 / 4:3 / 21:9 / 1:1. Top blind-test scores for visual quality (4.80/5) and text alignment (4.18/5).

Performance data (as of 2026-04-09)

| Track | HappyHorse-1.0 | Seedance 2.0 | Gap | |---|---|---|---| | T2V (no audio) | 1383 | 1273 | +110 | | I2V (no audio) | 1413 | ~1339 | +74 | | T2V (with audio) | ~1205 | ~1150 | +55 | | I2V (with audio) | ~1161 | ~1100 | +61 |

Win rates: 80.0% vs OVI 1.1; 60.9% vs LTX 2.3. Inference on one H100: 256p/5s ≈ 2s; 540p (with super-resolution) ≈ 8s; 1080p ≈ 38s.

Model scale

15B parameters, 40-layer Transformer (32 shared + 4 input + 4 output layers)
Runs on a single H100/A100 (48GB+ VRAM); weights ~30GB in FP16

Team and Alibaba's reorganization

Zhang Di joined Alibaba in 2010, was Kuaishou VP from 2020–2025 where he built Kling (dubbed the "Chinese Sora" at its June 2024 launch), briefly moved to Bilibili in 2025, and returned to Alibaba in November 2025. HappyHorse even surpasses his former employer's Kling 3.0 (Elo ~1241). Its launch coincided with major Alibaba restructuring: the ATH (Alibaba Token Hub) business group was formed on March 16, 2026 under CEO Wu Yongming, integrating Tongyi Lab, MaaS, Qwen, Wukong, and the AI Innovation unit (HappyHorse's home); a group technology committee followed on April 8.

Open source status and warnings

Alibaba claims full open sourcing — base weights, 8-step distilled model, super-resolution module, inference code, and commercial license — but as of April 10, 2026 the GitHub/Hugging Face repos still show "Coming Soon." Multiple fake websites (happyhorse.app, happy-horse.ai, etc.) have appeared; do not pay through unofficial channels. Official placeholder: https://happyhorsemodel.ai/ ; leaderboard: https://artificialanalysis.ai/video-arena

Competitive landscape

| Model | Vendor | Open source | Elo (T2V) | |---|---|---|---| | HappyHorse-1.0 | Alibaba | Claimed | 1333–1383 | | Seedance 2.0 | ByteDance | No | 1273 | | SkyReels V4 | Kunlun | Partial | 1245 | | Kling 3.0 | Kuaishou | Partial | 1241 | | PixVerse V6 | Aishi | Partial | 1241 |

Strengths: technical lead, native audio, multilingual lip sync, open weights, and reportedly half the price of Seedance. Weaknesses: no consumer product platform yet, 5–8s clips only, and a nascent community.

Industry implications

1. Open source wins: an open model topping the arena shows open communities have caught up with (or passed) closed leaders in video generation. 2. Joint audio-video is the future: "generate video first, dub later" workflows will become obsolete. 3. Chinese dominance in multimodal video: HappyHorse, Seedance, Kling, and SkyReels fill the leaderboard's top ranks.

Availability

Currently usable via Artificial Analysis blind voting; the Alibaba Bailian platform is in internal beta with public access expected within a week; official API and local deployment await the open-source release. Community quantized versions (FP8/INT8) may eventually bring it to consumer GPUs like the RTX 4090.