HappyHorse 1.1

Overview

HappyHorse 1.1 was released on June 23, 2026, the second version of a model line that Alibaba first shipped, in limited beta, in late April. Alibaba's announcement describes it as "the upgrade to its video generation model, to deliver enhanced creative quality, controllability, and production efficiency for professional content creators."

The model arrived by an unusual route. In early April 2026 an unidentified entrant called HappyHorse-1.0 appeared on the Artificial Analysis Video Arena and went straight to the top of the video generation rankings. Alibaba claimed it days later. Its own launch post, published April 28, 2026, credits an organisation most readers had never heard of: "Developed by Alibaba Token Hub (ATH) Business Unit."

That attribution matters, because ATH is not the Tongyi Lab that builds Wan. Chinese financial press reports that the team behind HappyHorse is the Future Life Lab (未来生活实验室), led by Zhang Di — who spent 2020 to 2025 at Kuaishou as a vice president and, per Sina Finance, was "技术一号位" (top technical leader) on Kling. Sina Finance reported his move to Alibaba on November 4, 2025, without giving a start date. If accurate, the man who architected Kling now competes with it. Alibaba has never confirmed this in an official HappyHorse announcement, and we mark it as press reporting rather than vendor fact.

On capability, Alibaba's own claim for 1.0 is that the model "supports generation of up to 15 seconds of 1080p video with multiple shots, and delivers synchronized audio-visual output — including lip-synced dialogue, ambient soundscapes, and emotionally expressive vocal performances."

A note on naming

Three corrections, because the web is unusually wrong about this model.

HappyHorse is not open source. Press releases syndicated to dozens of local-news sites announce "HappyHorse-1.0 Crowned #1 Open-Source AI Video Generator," and GitHub repositories call the model "100% open source" and offer to "Download full model weights." All of it is false. fal, Alibaba's official API partner, states that HappyHorse-1.0 "is closed source — it will not be open source or licensable." A Hugging Face API search for happyhorse returns zero models. arena.ai tags it Proprietary; Artificial Analysis's open-weights text-to-video sub-leaderboard is led by LTX, not HappyHorse. The GitHub repositories making the claim are individual user accounts, not the Alibaba organisation, and carry no license file at all.

ATH is not Tongyi, and HappyHorse is not Wan. arena.ai lists Alibaba-ATH and Alibaba as two different labs on the same leaderboard, with happyhorse-1.0 under the former and wan2.7-i2v under the latter. Alibaba Cloud Model Studio documents HappyHorse and 万相 (Wan) as separate families. Note one wrinkle: Alibaba's blog says ATH built it, while Sina Finance places the Future Life Lab in Taotian Group and describes the team as "此前为阿里ATH-AI创新事业部成员" ("previously members of Alibaba's ATH-AI innovation division"). The two accounts of the org chart do not perfectly agree, and we do not resolve them here.

Circulating Elo numbers are unreliable. Third-party write-ups quote HappyHorse's Artificial Analysis text-to-video Elo as 1333, 1357, 1374, and 1389, and its image-to-video Elo as 1392, 1402, and 1410. Arena Elo drifts continuously as votes accumulate, and Artificial Analysis and arena.ai use different scales. Every figure on this page is dated and labelled with its source.

Capabilities

• Four generation modes. Text-to-video, image-to-video (conditioned on a first frame), reference-to-video, and video editing — each a distinct Model Studio endpoint.
• Joint audio generation. Alibaba claims lip-synced dialogue, ambient soundscapes, and "emotionally expressive vocal performances" produced with the video, not dubbed on afterwards. fal reports native lip-sync across seven languages: Mandarin, Cantonese, English, Japanese, Korean, German, and French.
• Up to nine reference images. The reference-to-video endpoint accepts one to nine images; the prompt refers to them positionally as [Image 1], [Image 2], and so on, letting you pin a specific product, character, or scene into a generated shot.
• Video editing in two shapes. Alibaba names a Video-to-Video function that "allows users to modify an existing video while preserving its original structure and motion," and a Subject-and-Video-to-Video function that inserts or replaces a subject drawn from a reference image.
• Multi-shot sequencing. Alibaba claims stable character positioning across cut transitions within a single generation, which it pitches at short dramas.
• Any-language prompts. Prompts accept "any language," capped at 5,000 non-Chinese characters or 2,500 Chinese characters, with overflow silently truncated.

Technical Specifications

Alibaba publishes operational limits, and almost nothing about the model itself.

• Model IDs: happyhorse-1.1-t2v, happyhorse-1.1-i2v, happyhorse-1.1-r2v, and their 1.0 equivalents
• Video editing ID: happyhorse-1.0-video-edit — still at 1.0; no 1.1 editing model exists
• Resolution: 720P or 1080P, with 1080P the default
• Duration: an integer from 3 to 15 seconds, default 5
• Aspect ratios (text-to-video): 16:9 (default), 9:16, 1:1, 4:3, 3:4, 4:5, 5:4, 9:21, 21:9
• Watermark: true by default — a "Happy Horse" mark in the lower-right corner. Set watermark: false to remove it.
• Seed: integer in [0, 2147483647]; Alibaba warns a fixed seed "does not guarantee identical results"
• Reference images: 1–9 (reference-to-video); 0–5 plus exactly 1 video (video editing)
• Editing input limits: video 3–60 s, ≤100 MB, longer side ≤4,096 px, shorter side ≥360 px, aspect 1:2.5–2.5:1, >8 fps. Inputs beyond 15 s are truncated to their first 15 seconds.
• Invocation: asynchronous only (X-DashScope-Async: enable); task_id valid 24 hours; edits typically take 1–5 minutes
• Regions: China (Beijing), Singapore, US (Virginia), Germany (Frankfurt)
• Weights: none published. Parameter count: not disclosed. Alibaba publishes no architecture description, no parameter count, and no technical report, so this page does not state one. Widely repeated figures — 15 billion parameters, a 40-layer unified transformer, DMD-2 distillation to 8 steps — appear only in third-party blogs and trace to no primary source.

Note a naming inconsistency in Alibaba's own materials: the April launch post calls the third mode Subject-to-Video (S2V), while the June post and the API reference call it Reference-to-Video (R2V). They are the same endpoint.

Use Cases

Alibaba reports that since April, HappyHorse 1.0 "has been adopted across various content production sectors, including short-form dramas, e-commerce advertising, brand marketing, and gaming cinematic trailers (CG)." That list maps cleanly onto the model's design:

• Short-form drama — the multi-shot consistency and dialogue lip-sync target vertical serialised fiction, a large category in China.
• E-commerce advertising — reference-to-video pins an actual product's appearance into generated footage, which is the whole problem in commerce creative.
• Brand marketing — 21:9 and 9:21 aspect ratios alongside 9:16 cover cinematic and social framings from one model.
• Game cinematics — Alibaba explicitly names CG trailers.
• Post-hoc video editing — style transfer across an existing clip, or swapping a subject while preserving camera motion, via the editing endpoint.
• Localised dialogue — seven-language lip-sync, per fal, for one clip cut across markets.

Performance / Benchmarks

Two organisations rank HappyHorse. They use different Elo scales and their numbers must never be compared with each other. All figures below were read directly from the leaderboards on July 8, 2026 and will drift.

Artificial Analysis Video Arena

Text-to-video, without audio — the one track HappyHorse leads:

Text-to-video, with audio — a different story:

On image-to-video, HappyHorse leads neither track: HappyHorse-1.1 is 3rd with audio (Elo 1112) and 5th without (Elo 1317), behind Seedance 2.0 in both. Artificial Analysis states that models are "ranked using an Elo rating system derived from user votes in blind comparisons."

Two things follow. First, 1.0 outranks 1.1 without audio, and 1.1 outranks 1.0 with audio — consistent with Alibaba's claim that 1.1's work went into audio synchronization and motion, and a caution against assuming the newer model dominates. Second, HappyHorse's headline reputation rests on the track where its audio is switched off.

arena.ai (formerly LMArena)

arena.ai ranks only happyhorse-1.0; happyhorse-1.1 is in the voting pool but not yet ranked.

On both boards the lab is listed as Alibaba-ATH · Proprietary, distinct from the Alibaba entries for the Wan models.

Everything above is a blind-vote preference score, not a measurement of fidelity, physics, or instruction following. Alibaba's own quality claims — "smoother, more coherent, and visually striking movement," "richer visual details and lifelike imagery" — are vendor claims, published without benchmark numbers of any kind.

Limitations

• Vendor claims are unquantified. Alibaba published no benchmark table, no ablation, no technical report. Every 1.1-versus-1.0 improvement is asserted in prose.
• No weights, no self-hosting, no fine-tuning. Closed model, API only. For open weights in this category see Wan 2.2, which is Apache-2.0.
• Audio is undocumented in the API. Alibaba markets synchronized audio, and Artificial Analysis ranks the model in "with audio" tracks — yet the Model Studio API reference for text-to-video and image-to-video, in both English and Chinese, contains no audio parameter and no mention of audio at all. On those two endpoints there is no documented way to disable it, control it, or select a language. Only the video-editing endpoint exposes an audio_setting parameter.
• Watermarked by default. Generations carry a "Happy Horse" mark unless you explicitly opt out.
• Fifteen seconds, hard. No long-form generation; editing inputs beyond 15 s are silently truncated.
• The editing endpoint is a version behind and bills for input and output seconds, so a 15-second edit of a 15-second clip is charged as 30 seconds.
• Singapore costs materially more — roughly 17% above the mainland rate at 720P — for identical output.
• Ranking is narrower than advertised. It leads exactly one of Artificial Analysis's four video tracks, and is 4th on both arena.ai boards.

Pricing & Access

Alibaba publishes HappyHorse pricing in yuan only, per second of generated video, on Model Studio's model pricing page. Rates below are reproduced in CNY; we do not convert. Billing is output-only (仅输出计费) for generation: 费用 = 视频单价 × 输出的视频时长 — "cost = unit price × output video duration." Failed requests are not billed and do not consume free quota.

Generation — text-to-video, image-to-video, reference-to-video

1.1 is 25% cheaper than 1.0 at 1080P and identical at 720P.

Video editing

happyhorse-1.0-video-edit is ¥0.9/second at 720P and ¥1.6/second at 1080P (mainland China, US, Germany). Its billing rule differs: 输入视频和输出视频均计费 — both the input and the output video are billed by the second.

Free quota

10 seconds of generation per model, valid for 90 days after activating Model Studio. Mainland China deployment only; Alibaba states other regions receive no free quota.

Access routes

• happyhorse.com — the official product site, named as such in Alibaba's own 1.1 announcement.
• Alibaba Cloud Model Studio (百炼 / Bailian) — the API, and the only route with published rate cards.
• Qwen App — Alibaba's consumer AI application; named in the 1.0 launch post as the individual-user surface.
• fal — "official API partner," serving HappyHorse 1.0 across all four endpoints at $0.14/second (720p) and $0.28/second (1080p). fal publishes no 1.1 endpoint.

Ecosystem & Tools

• HappyHorse API reference - the hub for all four endpoints
• Text-to-video reference - resolution, ratio, duration, watermark, seed
• Reference-to-video reference - the 1-to-9 reference-image rules
• Video editing reference - input video constraints and the dual-billing rule
• Model Studio pricing - the CNY rate card, by region
• fal - USD per-second pricing and the only English capability sheet
• HorsePower AI Video Competition - announced alongside 1.1 with Huajing Entertainment Group; TechNode reports commercial collaborations valued up to RMB 1 million

Community & Resources

• Alibaba Rolls Out HappyHorse 1.0 in Limited Beta - the April 28, 2026 launch post, and the source of the ATH attribution
• HappyHorse Gets Stronger Motion Expressiveness… - Alibaba's June 23, 2026 announcement of 1.1
• Artificial Analysis Text to Video Leaderboard - the no-audio track HappyHorse-1.0 leads
• arena.ai Image-to-Video Leaderboard - Alibaba-ATH · Proprietary, rank 4
• Sina Finance, November 4, 2025 - Zhang Di leaves Kuaishou for Alibaba's Future Life Lab
• Sina Finance, April 9, 2026 - Chinese press identifies the anonymous chart-topper
• TechNode on HappyHorse 1.1 - the 1.1 launch and the filmmaking competition

Compare with Kling 3.0 — the model Zhang Di is reported to have architected before this one — and with Seedance 2.0, which outranks HappyHorse on every track where audio counts. See also Wan 2.2 for Alibaba's other, open-weight video line, plus Sora 2, Hailuo 2.3, and Vidu Q3 Pro.