GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —
GPT-56T 861 —
MUSE-SPK 837 +0.2%
GPT-56SC 790 -4.6%
GLM-5 781 -0.4%
CL-OP55X 780 -5.1%
GROK-46H 780 -5.1%
QWEN-38X 748 -9.2%
GPT-6A 743 -9.4%
KIMI-K3X 742 —
CL-FAB5H 698 -6.1%
CL-OP5H 675 -6.2%
GEM-38FH 672 -0.7%
CL-OP5X 670 -5.5%
CL-OP55H 668 —
CL-OP46H 657 -5.9%
CL-OP47H 648 -6.1%
GPT-56S 618 -0.6%
GEM-37FH 610 -7.2%
GEM-36FH 593 —
CL-OP48H 588 —
CL-OP47 581 -0.2%
GEM-35FH 580 —
GPT-55H 541 -7%
INKL 531 —
GEM-31P 512 -0.2%
CL-OP46 498 +0.4%
GEM-3P 498 -0.2%
CL-OP48 492 +0.4%
GPT-52 464 —
GPT-55 423 —
← Back to feed

Alibaba's HappyHorse-1.0 Tops Artificial Analysis Video Arena at Elo 1389 T2V, 1411 I2V

A 15-billion-parameter video model from inside Alibaba’s e-commerce arm has taken the top position on the Artificial Analysis Video Arena across two categories, dethroning ByteDance’s SeedDance 2.0 by a margin that makes the prior leader look like a placeholder.

HappyHorse-1.0 holds Elo 1389 in text-to-video (no audio) and Elo 1411 in image-to-video (no audio). SeedDance 2.0, which held both spots before HappyHorse arrived in late March, sits at roughly Elo 1274 T2V — a 115-point gap on a platform where 10 points is a meaningful lead. In the with-audio categories, HappyHorse ranks first or second, competitive with SeedDance 2.0 to within a few points.

Architecture

The model is a single unified 40-layer Transformer that jointly denoises video tokens and audio tokens in one pass. No separate audio pipeline, no modality alignment step. This is the same approach that ByteDance pursued in SeedDance but executed at a higher parameter count with what appears to be better training discipline.

Key specs:

  • 15B total parameters, 40-layer unified self-attention Transformer
  • Native 1080p output at 24fps
  • Synchronized multilingual audio: Chinese, English, Japanese, Korean, German, French
  • Inference: 38.4 seconds for a 5-second 1080p clip on a single H100
  • 8-step DMD-2 distillation for fast inference
  • 7x KV-cache reduction for long-context efficiency
  • Apache 2.0 license

Who built it

The lab behind HappyHorse is Future Life Lab inside Taotian Group’s Alibaba Token Hub (ATH) innovation unit. The lab is led by Zhang Di, a former VP at Kuaishou who was the technical lead on Kling AI — the video model that itself became the benchmark-topper in 2024 before Wan and SeedDance caught up. Zhang Di’s team joined Alibaba in late 2025.

The situation: the person who built Kling is now running the lab that has produced a model that beats Kling 3.0 on both major video leaderboard tracks.

How it surfaced

HappyHorse first appeared anonymously on the Artificial Analysis Video Arena around April 7, 2026 — no company name, no paper, no landing page. It climbed to number one within days. Alibaba confirmed ownership on April 10 via the official @AlibabaGroup account. The model has held its rankings since.

Developer API access opened April 27 via fal, which is among the first official API providers. Pricing and rate limits have not been disclosed.

Benchmark context

The Artificial Analysis Video Arena uses blind Elo evaluations: users compare two unlabeled clips and vote, with no knowledge of which model produced either. HappyHorse’s current T2V ranking is based on over 13,900 community samples. This methodology is harder to game than self-reported benchmark suites.

The model leads over Kling 3.0 Pro, PixVerse V6, and every other entrant on the no-audio tracks. On with-audio tracks it is within error margin of SeedDance 2.0 720p.