HappyHorse-1.0 appeared in the last 48 hours like a ghost in the machine, and by Tuesday morning it sat atop the Artificial Analysis Video Arena leaderboard with an Elo score between 1,367 and 1,402 across text-to-video and image-to-video categories — outpacing Seedance 2.0, Kling 3.0, and every other closed model in the field. Eight thousand blind human votes don't lie. The model won 80% of head-to-head matchups against Ovi 1.1 and nearly 61% against LTX 2.3. Its Visual Quality score hit 4.80. Physical Consistency reached 4.52. Lip-sync accuracy stretched across seven languages — Mandarin, Cantonese, English, Japanese, Korean, German, French — with a Word Error Rate of just 14.60%.
Then, just as abruptly as it appeared, some versions vanished. The GitHub and Hugging Face links read: Coming Soon.
The Decryption
The AI community scrambled. Speculation ran from Alibaba's WAN video family to obscure research labs. By Wednesday, sources converged on an answer that reframed everything: HappyHorse-1.0 is the work of Alibaba's Taotian Group — the e-commerce arm behind Taobao and Tmall — operating through an internal unit called the Future Life Lab, housed within the ATH-AI Innovation Division. Leading the effort is Zhang Di, formerly Vice President at Kuaishou and the technical architect behind its Kling AI video model. He joined Alibaba at the close of 2025 to spearhead multimodal AI development.
The reveal coincided with notable movement in Alibaba's Hong Kong stock.
Technically, the model is a marvel of restraint. Fifteen billion parameters. A single-stream 40-layer Transformer with no cross-attention — text, video latents, and audio tokens unified into one sequence. A "sandwich" architecture: four modality-specific outer layers wrapped around 32 fully shared inner layers. DMD-2 distillation compresses generation to just eight denoising steps, eliminating the need for classifier-free guidance entirely. On a single H100 GPU, it renders five seconds of 1080p footage in 38.4 seconds. A custom compiler called MagiCompiler adds another 1.2× speedup on top.
The official site calls it an "Open Release" under commercial-friendly licensing, with a rumored full open-source drop — GitHub, Hugging Face, inference code, base and distilled models — potentially arriving as soon as April 10, 2026.
What the Benchmarks Don't Capture
"The biggest issues with current top video generation models are availability, pricing, and excessive restrictions."
That is the conclusion reached by the engineering team at CTOL Digital Solutions, which conducted independent head-to-head evaluations of HappyHorse-1.0 against Seedance 2.0 across real client production cases — the kind of testing no leaderboard can replicate.
Their findings were measured and pointed: Seedance 2.0 still edges ahead in naturalness, physical logic, and camera motion handling. With respect to motion control, prompt adherence, and realistic detail in complex real-world scenes, the gaps were both obvious and nontrivial — and stubbornly invisible to benchmark scores. HappyHorse, they observed, lacks a certain liveliness, a human warmth that Seedance carries into its more demanding outputs.
"I'm honestly skeptical of anyone claiming HappyHorse has already surpassed Seedance 2.0," the CTOL team wrote. "I don't believe it just yet."
But their conclusion pivots sharply. Both Seedance and Google's Veo are hamstrung by aggressive content protection — rendering them "unusable for anything with even minimal copyright risk.." Veo, they note, is the worst offender. Moreover, following Sora's unexpected retreat, Seedance and Veo remain largely unavailable and come with high price tags. HappyHorse, if it delivers on its open-source promise, sidesteps those walls entirely.
"If it can deliver something better in availability, price, and restrictions through open source, it will be the clear winner."
This is the pattern of Chinese AI in 2026: anonymous drops, rapid community decryption, and models that arrive not with fanfare but with receipts. The Horse Year, it seems, is running fast.
Sources: https://x.com/jiqizhixin/status/2041814095977181435