Alibaba's "HappyHorse" Gets Faster: HappyHorse 1.1 Officially Released

On June 22, Alibaba officially released HappyHorse 1.1, its native multimodal AI video generation model. Less than two months after the 1.0 gray launch, this update targets real production pain points: stiff motion, face drift, prompts that get ignored, and over-sharpened “oily” frames.

If you’ve been searching for a happyhorse tutorial or comparing happyhorse video tools, this guide walks through what 1.1 actually changed.

Where 1.0 left off

Since HappyHorse 1.0 opened for gray testing on April 27, the model has been adopted across short drama, e-commerce ads, brand marketing, and game CG pipelines.

The core architecture stays the same: a 15-billion-parameter single-stream Transformer that encodes text, image, video, and audio in one sequence for native audio-video generation. Version 1.0 topped Artificial Analysis Video Arena text-to-video and image-to-video rankings — a high bar for 1.1 to beat.

What 1.1 upgrades: five dimensions

Dimension	Key changes
Dynamic expression	Better motion modeling and temporal consistency
Subject consistency	Up to 9 reference images at once; stronger R2V fusion
Instruction following	Long-context understanding; multi-shot and N-grid references
Visual quality	Less oily look and over-sharpening; more skin texture detail
Audio expression	Improved native joint audio-video output

Motion: less “stop-motion” feel

Stiff character motion and disconnected camera rhythm were among the most common 1.0 complaints. Version 1.1 improves motion modeling and frame-to-frame coherence — runs, jumps, and object collisions feel more physically plausible.

For happyhorse video workflows, that means fewer re-rolls just to get usable action.

Visual fidelity: less plastic skin

Creators often noted 1.0 frames could look overly smooth or harshly sharpened. Version 1.1 addresses this directly:

• Reduced oily sheen and excessive sharpening
• Preserved pores, fine lines, and natural skin texture
• Better micro-details: sweat, lighting shifts, debris on impact

Short drama and ad workflows benefit most from these changes.

Multi-image R2V: lock identity with up to 9 refs

Reference-to-Video (R2V) is a headline upgrade. Upload up to 9 reference images and tag them in your prompt as character1, character2, etc. The model fuses identity, wardrobe, and brand elements into a coherent clip.

Use case	What 1.1 R2V delivers
Product ads	Stable logos and packaging details
Short drama	Same character across multiple shots
Multi-shot scripts	Better N-grid reference parsing and scene planning

Prompts that actually stick

Long prompts with multiple scenes and characters used to confuse the model. Version 1.1 improves instruction following and shot planning — the output tracks your intent more reliably.

Where to access it

HappyHorse 1.1 is live on the official site, Alibaba Cloud Model Studio (Bailian), and the Qwen app. Enterprise teams and developers can call the full capability set via API.

Try it yourself

Ready to test HappyHorse 1.1 happyhorse video generation? Use the button below to open the official app and pick the video generation tool.

New users typically get free trial credits — run the same prompt on 1.0 and 1.1 to compare.

Bottom line

HappyHorse 1.1 is a production-driven upgrade, not a vanity version bump. Smoother motion, stabler faces, and stronger multi-image R2V make it worth a side-by-side test if you’re already on 1.0 — and a good entry point if you’re new.