Alibaba HappyHorse 1.1 Goes Live: This Time the Open-Source Video Model Fixed the Oily Gloss Issue

On June 22, Alibaba launched HappyHorse 1.1. The HappyHorse official website, Alibaba Cloud Bailian, and Qianwen Cloud all integrated it simultaneously. Existing users don’t need to change their workflows—just tweak the version number.

What makes this worth mentioning is that HappyHorse 1.0 only last month topped the Artificial Analysis Video Arena leaderboard with an Elo score of 1333, boasting an 80% win rate against OVI 1.1 and a 60.9% win rate against LTX 2.3—currently the highest-ranked open-source video generation model in the world. Before 1.0 cooled off, 1.1 already arrived, with an iteration pace clearly faster than models from typical large companies.

First, What Hasn’t Changed: Specs Are Still the Same

HappyHorse 1.1 retains the same underlying specs as 1.0:

• Single generation length: 3 to 15 seconds
• Resolution: 720p / 1080p
• Aspect ratio: Flexible
• Number of parameters: 15 billion (same architecture as 1.0)
• Structure: 40-layer unified self-attention Transformer
• Audio-video: Native joint generation, lip-sync in seven languages (English, Mandarin, Cantonese, Japanese, Korean, German, French)

That 1.0 benchmark of “Single H100 generating 5s 1080p video in just 38 seconds” still has few rivals in the open-source camp. This means 1.1 wasn’t retrained from a new base, but fine-tuned and patched on the existing architecture. From an engineering standpoint, this kind of iteration costs less, poses lower regression risk, and is more aligned with commercial pace—preserving existing pipelines so enterprise clients can switch smoothly.

Five Upgrade Dimensions – Which Are Truly Useful

The official upgrade list covers these five aspects: dynamic expressiveness, subject consistency, instruction adherence, visual texture, and audio capability. Let’s examine them one by one.

1. Dynamic Expressiveness: Solving the “PPT Feel”

Anyone who’s worked on video generation optimization knows movement modeling and temporal consistency are somewhat at odds—if you want steady visuals, the model tends to become “lazy,” with slower and smaller movements, resembling an advanced photo animation. In complex motion scenes, 1.0 suffered from this.

1.1 specifically optimized this area. In the official words: “Make movements of people and objects more natural and coherent, with greater visual tension and sense of force.” Testing dance, action, and sports scenes shows noticeably less “floating” compared to 1.0, and tighter transitions between keyframes. For short videos, ads, and cinematic storyboards—these high-motion-demand scenarios—this is a real improvement, not just a demo stat.

2. Subject Consistency: A Commercial Must-Have

This is an enhancement to its R2V (Reference-to-Video, multi-image→video) capability, arguably the upgrade most aligned with real commercial use.

Specific problems addressed:

• Retaining precise product details and brand elements (critical in e-commerce ads)
• Flexibly combining characters and scenes while keeping subject stable (essential for narrative content)
• Correctly interpreting multi-shot and N-grid references (common in ads and music videos)

Previously, open-source video models performed mediocrely here. With many references or complex combinations, models would start “guessing”—logos distorted, faces blurred, scenes misaligned. In 1.1, improvements in multi-source reference semantic understanding and fusion are evident. Product textures and brand detail fidelity are visibly better in official examples.

This clearly signals intent to compete head-on in advertising and e-commerce sectors.

3. Instruction Adherence: Long Prompts No Longer “Crash”

Complex narrative prompts have always been a challenge for video models—long text, multiple scenes, intricate character relationships often result in missing elements, confusion, or compressing multi-segment descriptions into a single frame.

1.1 strengthens three capabilities:

• Long-context semantic understanding: Can digest longer, more complex prompts
• Scene planning: Knowing which description maps to which shot
• Character relationship modeling: Understanding who’s doing what and their relationships without confusion

For workflow developers, this is crucial. In AIGC production chains, copywriting → storyboard → generation → editing is linear; no matter how well you engineer prompts, if the model’s comprehension is lacking, results suffer. 1.1 aims to give the model some “narrative arrangement” capability, essentially shifting work that used to require human engineers into the model itself.

4. Visual Texture: Finally Fixed the Oily Gloss and Smearing

This might be the most demanded improvement by the developer community.

AI-generated human visuals have long suffered from:

• Over-sharpening, making skin look like reflective paper
• Oily gloss, as if everyone just came from the gym
• Smearing, excessive smoothing erasing real detail
• Detail instability—either absent or monstrously exaggerated

1.1 optimizes facial detail generation, authentic skin reproduction, and camera language understanding, explicitly aiming to “preserve real features like acne marks, nasolabial folds, and pores while avoiding detail over-amplification.” This wording is noteworthy—it acknowledges the past struggle with balancing “realism” and “excess,” and makes subtlety a target.

Comparatively, overseas Seedance 2.0 and Keling 3.0 each have their own texture preferences; HappyHorse 1.1 clearly aims for “cinematic” and “ad-grade realism.”

5. Audio Capability: Continuation of Native Audio-Video Joint Generation

The official note doesn’t elaborate, but HappyHorse 1.0 was the world’s first open-source large video model to natively support audio-video joint generation, with the lowest lip-sync word error rate among its peers. 1.1 continues to optimize here, maintaining lip-sync ability across seven languages.

Notably, this “native joint” approach differs from “generate video first, then post-process audio” methods. The latter stitches two models, capping alignment precision; the former treats audio and video as a unified modality from training, with a theoretically higher ceiling.

Horizontal Comparison: How It Stacks Against Seedance, Keling, OVI, LTX

Based on 1.0 benchmarks:

| Comparison Dimension | HappyHorse 1.0 vs Competitors | |----------------------|------------------------------| | Text→Video (no audio) | Beats Seedance 2.0, Keling 3.0 | | Image→Video (no audio) | Beats Seedance 2.0, Keling 3.0 | | Text→Video (with audio) | Slight lead | | Image→Video (with audio) | Ties Seedance 2.0 | | vs OVI 1.1 | 80% win rate | | vs LTX 2.3 | 60.9% win rate |

The 1.1 upgrades clearly target the “Image→Video (with audio)” weak spot—subject consistency and texture improvements directly correspond to this path. If new blind tests roll out, this tie could turn into a lead.

Zhang Di’s Team + Zhang Jizhong + Million-Dollar Contracts: Alibaba’s Video Ambition

HappyHorse is developed by a team led by former Kuaishou VP Zhang Di—obviously bringing short-video expertise to Alibaba’s generative video efforts.

This release also announced the “Horsepower” AI Video Contest co-hosted with Orca Media Group, judged by Zhang Jizhong. Noteworthy points:

1. Traditional film industry endorsement: Zhang Jizhong represents recognition from traditional content production, key for AIGC content entering mainstream film.
2. Million-dollar commercial cooperation: Awarding winners real commercial deals—targeting professional creators, not hobbyists.
3. Ecosystem loop: Model + contest + contract distribution—similar to Kuaishou’s past creator support strategy.

Open-source access + commercial ecosystem loop is Alibaba’s standard approach in the large-model field. Comparing with Tongyi Qianwen’s method reveals the same methodology applied to video.

What This Means for Developers

Practical takeaways:

• Open-source truly means open: 15-billion parameter, 40-layer Transformer architecture details were public in 1.0, and 1.1 will likely be too. This enables private deployment and secondary fine-tuning.
• Low inference cost: Single H100 generates 5s 1080p video in only 38 seconds, impressive for a 15B parameter video model.
• R2V capabilities directly usable in e-commerce: 1.1’s subject consistency boosts SKU video generation effectiveness.
• Complete API integration paths: Available in Alibaba Cloud Bailian and Qianwen Cloud—enterprise clients can go cloud, personal developers can use the official site.

If your product compares multiple video models, create a benchmark: run the same prompt on HappyHorse 1.1, Seedance 2.0, and Keling 3.0, then examine subject consistency and texture. In 1.0 the differences were subtle; post-1.1, you may see a clear shift.

A Final Assessment

Video generation model iteration pace now resembles 2023’s language models—major versions every two to three months, each with visible ability upgrades. HappyHorse’s leap from 1.0 to 1.1 in under two months shows it’s keeping up.

More importantly, 1.1’s focus isn’t parameter scaling, long video ambition, or flashy effects—it’s addressing concrete commercial pain points: subject consistency, realistic texture, and long prompt comprehension. This indicates Alibaba’s clear positioning for the model: it’s not for leaderboard glory—it’s for making money.

Few in the open-source camp take this approach. Worth watching closely.

OpenAI Hub (openai-hub.com) currently supports aggregated calls for mainstream video generation models. HappyHorse series integration is planned—those needing multi-model comparisons should take note.

References

• ITHome: Alibaba Releases Video Generation Model HappyHorse 1.1 — First-hand report of official upgrade details and capability descriptions