HappyHorse-10 Deep Dive: Alibaba's Surprise Open-Source Model Tops AI Video Arena

Forum topic · 小凯 · 2026-04-10

Summary

HappyHorse-1.0, an anonymously released video generation model that traces back to Alibaba Taotian Group's Future Life Lab, stunned the industry by topping the Artificial Analysis Video Arena on April 7, 2026, with an Elo score of 1333, roughly 60 points ahead of ByteDance's Seedance 2.0, and ranking first in all four tracks (text-to-video and image-to-video, with and without audio). Built as a 15B-parameter, 40-layer single-stream Transformer (32 shared layers, 4 input and 4 output layers), it jointly generates video and audio in one pass, eliminating separate dubbing and lip-sync pipelines. Key techniques include DMD-2 distribution-matching distillation (8 denoising steps, no classifier-free guidance), per-head sigmoid gating for stable multimodal training, and timestep-free denoising. It generates 5-second 1080p video in about 38 seconds on one H100, supports 7-language lip sync, and is claimed fully open source with commercial licensing, though weights were still marked "Coming Soon" as of April 10, 2026. The team is led by Zhang Di, former Kuaishou VP known as the father of Kling. The model's lead suggests architecture innovation and open-source efforts can outperform larger closed models in video generation.

HappyHorse-1.0 Deep Dive: Alibaba's "Happy Horse" Reshapes AI Video Generation Overnight

*Structured summary of a Chinese forum post (zhichai.net). All figures are as reported by the original author.*

Key points

On April 7, 2026, a model called HappyHorse-1.0 appeared anonymously on the Artificial Analysis Video Arena and topped the leaderboard with Elo 1333, about 60 points ahead of ByteDance's Seedance 2.0.
It ranked first in all four tracks: T2V and I2V, each with and without audio. As of April 9, it scored 1383 in pure video quality (T2V, no audio), 110 points ahead of 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 (the "father of Kling"), who returned to Alibaba in November 2025 after a brief stint at Bilibili.
Head-to-head win rates: 80.0% vs OVI 1.1, 60.9% vs LTX 2.3.

Technical architecture

HappyHorse's core idea is unified, end-to-end generation — video and audio are produced together in a single forward pass, not stitched together in post.

Single-stream Transformer (40 layers):
4 input layers embed text, image, video, and audio tokens
32 shared layers process all modalities in one sequence via self-attention (no cross-attention modules)
4 output layers decode video frames and audio waveforms
DMD-2 distillation (Distribution Matching Distillation): compresses denoising from 50+ steps to 8 steps and removes the need for Classifier-Free Guidance (CFG), halving per-frame inference work.
Per-Head Sigmoid Gating: a learnable scalar gate per attention head (sigmoid(gate) * attention(x)) selectively suppresses conflicting gradients between modalities, stabilizing joint training.
Timestep-free denoising: the model infers noise level from the latent itself, simplifying the architecture and enabling extreme distillation.
With the internal MagiCompiler runtime, it generates 5s 1080p video in ~38 seconds on a single H100 (256p preview: ~2s; 540p with super-resolution: ~8s).

Capabilities

Native joint audio-video generation: lip-synced speech, ambient sound (footsteps, wind, rain), and Foley effects — semantically aligned with the visuals (e.g., footsteps match walking rhythm).
7-language lip sync: English, Mandarin, Cantonese, Japanese, Korean, German, French; reportedly the lowest word error rate among comparable open models.
Output specs: native 1080p, 5–8 seconds (extendable), aspect ratios 16:9 / 9:16 / 4:3 / 21:9 / 1:1, standard cinematic frame rates. Blind-test scores: 4.80/5 visual quality, 4.18/5 text alignment.
Model size: 15B parameters — deployable on a single H100/A100 (48GB+ VRAM); weights ~30GB (FP16).

Benchmark standings (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 |

Team and organizational context

Zhang Di's trajectory: joined Alibaba in 2010; Kuaishou VP 2020–2025 where he built Kling's technical foundation (Kling 1.0 launched June 2024, dubbed the "Chinese Sora"); brief move to Bilibili in 2025; returned to Alibaba in November 2025 to head the Future Life Lab; released HappyHorse in April 2026 — reportedly surpassing his former employer's latest Kling 3.0.
March 16, 2026: Alibaba formed the ATH business group (Alibaba Token Hub) under CEO Wu Yongming, consolidating Tongyi Lab, MaaS/Bailian, Qwen (C-end), Wukong (B-end), and the AI Innovation unit housing HappyHorse.
April 8, 2026: a group technology committee was formed, also led by Wu Yongming; Tongyi Lab was upgraded to the Tongyi Large Model division.

Open-source status and caveats

Officially claimed as fully open source (base weights, 8-step distilled model, super-resolution module, inference code, commercial license), but GitHub and Hugging Face repos still showed "Coming Soon" as of April 10, 2026.
Scam warning: fake official sites (e.g., happyhorse.app, happy-horse.ai) have appeared; do not pay through unofficial channels before weights are released.
Reported weaknesses: no consumer product platform yet (unlike Seedance's Jimeng or Kling's ecosystem), 5–8s clip length limits, and a still-small community.

Industry implications

1. Open source wins: the first open model to top the video arena, showing open models can match or beat closed leaders in video generation. 2. Joint audio-video generation is validated: the "generate silent video, then dub" workflow may become obsolete. 3. Chinese models lead multimodal video: HappyHorse, Seedance, Kling, and SkyReels dominate the top ranks, even as the US leads in LLMs. 4. Architecture over scale: 15B parameters beating larger closed models underscores the value of design innovation.

How to access

Artificial Analysis arena: available now for blind-test voting — https://artificialanalysis.ai/video-arena
Alibaba Bailian platform: internal beta, expected public opening within about a week
Official API and local deployment: pending official release/weight drop
Official placeholder page: https://happyhorsemodel.ai/

Expected local requirements: NVIDIA H100/A100 (48GB+ VRAM), 64GB+ RAM, ~30GB storage (FP16), Python 3.10+, PyTorch 2.0+. Community FP8/INT8 quantizations may lower the bar to consumer GPUs like the RTX 4090.

Outlook

Expect the official weight release, longer 15–30 second versions, deep integration with Alibaba's e-commerce scenarios (product videos, livestream clips), and competing models adopting joint audio-video architectures in the coming months.