HappyHorse 1.1 Guide: What Alibaba Changed and What It Costs

Alibaba's video model quietly became a problem for everyone else. HappyHorse 1.1 landed with a claim that would have sounded absurd a year ago: production-ready video synthesis, third place in the public arena conversation and — in VentureBeat's framing — rising to No. 2 in global rankings while OpenAI's Sora and ByteDance's Seedance lose ground.

For anyone building a video workflow, that matters for a boring reason. Every time a top model gets cheaper and more consistent, the price of the whole category moves. HappyHorse 1.1 is now at roughly $0.145 per second at 1080p through a per-second API — the same band as Wan 3.0, and a fraction of what comparable long-clip models charge.

This guide covers what actually changed from 1.0, what the limits are in practice, what it costs per finished clip, and how to try it without committing to anything.

TL;DR

• HappyHorse 1.1 is Alibaba's upgrade to a 15B-parameter video model with native audio, handling text-to-video, image-to-video and reference-to-video in one pipeline.
• The real gains are consistency and instruction-following — character identity across shots, lip-sync stability, and better behaviour on long, detailed prompts.
• Clips run 3–15 seconds at 720p or 1080p, 24fps, with 9:21 through 21:9 aspect ratios and up to nine reference images.
• Cost is the headline: about $0.11/second at 720p and $0.145/second at 1080p, so a 15-second 1080p clip is roughly $2.18.
• It does not do 30-second single passes. If that is your requirement, Seedance 2.5 is still the only option — at roughly four times the per-second price.

Generate a Free Baseline Clip on Wan →

What HappyHorse 1.1 is

HappyHorse is Alibaba's joint audio-video generation model. The 1.0 release established the architecture: a 15-billion-parameter model built on a 40-layer unified self-attention transformer with DMD-2 distillation, which cuts denoising to about eight steps and generates 1080p output in roughly 38 seconds. It shipped with native joint audio-video synthesis and lip-sync across seven languages.

The 1.1 upgrade does not replace that architecture — it fixes the parts that made 1.0 feel like a demo:

Area	HappyHorse 1.0	HappyHorse 1.1
Prompt understanding	Good on short prompts	Better on long, multi-beat prompts
Character consistency	Drifted across shots	Reference-based identity holds
Motion	Convincing, occasionally floaty	Smoother, better weight and contact
Lip-sync	Multilingual, some drift	Positioned as zero-drift, with context-aware pacing
Audio	Dialogue-focused	Dialogue plus ambience, music and Foley in one pass
Duration	Shorter clips	3–15 seconds per generation

The audio point is the one people underrate. Most video models treat sound as a second stage — generate video, then add voice, then add effects. HappyHorse generates them together, so a door slam and the dialogue that follows it share timing instead of being nudged into place in an editor.

The three input modes, and when to use each

Mode	Input	Best for	Watch out for
Text to video (T2V)	Prompt only	Concepts, b-roll, mood pieces	Least control over framing
Image to video (I2V)	One still	Product shots, portraits, existing artwork	Motion is inferred from a single frame
Reference to video (R2V)	Up to 9 images	Character consistency, brand assets, multi-shot	Bad references produce confident garbage

R2V is where 1.1 earns its keep. Up to nine reference images, each at least 400px on the shortest side, up to 20MB, in JPEG, PNG or WebP. That is enough for a character turnaround plus a product plus a style plate — the three things that usually drift.

In practice: T2V for exploring, I2V for animating an approved frame, R2V for anything that has to look the same in shot two as it did in shot one.

Specs and limits in one table

Spec	Value
Resolutions	720p, 1080p
Frame rate	24 fps
Duration	3–15 seconds, whole seconds
Aspect ratios	16:9, 9:16, 3:4, 4:3, 4:5, 5:4, 1:1, 9:21, 21:9
Reference images	Up to 9 (JPEG, PNG, WebP; ≥400px shortest side; ≤20MB each)
Audio	Native, generated in the same pass
Lip-sync	Multilingual
Enterprise access	Alibaba Cloud Model Studio, with SLAs and compliance

Two of these deserve a second look. First, the aspect ratio list is unusually complete — 4:5 and 5:4 matter for feed placements, and 9:21 covers vertical stories without cropping. Second, 15 seconds is a hard ceiling: longer pieces need stitching or a different model.

If you want output today, start here: Launch Wan 2.7 Now →

What it actually costs

Per-second pricing makes the maths trivial, which is exactly why it is worth doing.

Resolution	Rate	5s clip	10s clip	15s clip
720p	~$0.1125/s	$0.56	$1.13	$1.69
1080p	~$0.145/s	$0.73	$1.45	$2.18

For context, a 15-second 1080p clip costs about $2.18 here, about $2.10 on Wan 3.0, and about $8.85 on Seedance 2.5 at gateway pricing. The gap between HappyHorse and Wan is noise. The gap between either of them and Seedance is the cost of a 30-second single pass.

What that means for planning: budget by retries, not by clip count. Three takes per usable clip is normal, so a 15-second finished shot at 1080p is realistically $6.50, not $2.18. Still cheap — but cheap enough that you will iterate more, which is the actual behaviour change.

Try Free AI Video Generation First →

How to prompt it

HappyHorse follows scene direction closely, which is both the opportunity and the trap: vague prompts get you generic footage, and detailed prompts get followed literally, including the parts you did not mean.

The structure that works:

[Shot type and subject] — [action with a clear start and end] — [camera behaviour] — [lighting and time of day] — [audio you want in the scene] — [what must not change]

Three examples:

Product spot

Medium close-up of a matte black water bottle rotating slowly on a stone plinth at golden hour, single slow push-in, hard sunlight from camera left with a long shadow, no people, no text. Audio: low room tone, a faint metallic click at the end.

Character dialogue

Two-shot of a courier in a wet navy jacket handing a parcel to a woman in a doorway at night, rain visible in the streetlight, handheld camera with slight drift, warm interior light spilling onto the step. She speaks one short line, he nods, then turns to leave. Audio: rain, footsteps, the door chain.

Reference-led brand scene

Using @Image1 for the character's face and jacket and @Image2 for the packaging: the character places the box on a counter and slides it forward, medium shot, soft window light from the right, shallow depth of field, camera locked off. Audio: cardboard on wood, quiet room tone.

Note the audio clause. Most people write prompts as if sound is added later. Here it is generated with the picture, so name it.

And the negative rule: never write "no distortion", "no extra fingers", "no warping". State what should be true — "one pair of hands, five fingers visible, product label facing camera" — and the model has something to build toward.

Ready to try it yourself? Try Wan 2.7 Free →

Test the Same Prompt on Wan, Free →

HappyHorse 1.1 vs Wan: which to use

Both are cheap and both are good. The split is about duration and control surface, not quality.

Dimension	HappyHorse 1.1	Wan 2.7
Duration per generation	3–15s	2–15s
Native audio	Yes, generated with video	Yes, with voice cloning
Reference images	Up to 9	Reference video and image support
First/last frame control	Limited	Yes — a core feature
Lip-sync	Multilingual, zero-drift claim	Clone a voice sample, sync to it
Aspect ratios	9 options including 4:5 and 9:21	Standard set
Price at 1080p	~$0.145/s	Subscription or per-second front ends

Choose HappyHorse 1.1 when the scene needs believable ambience and dialogue in one pass, when you want an unusual ratio without cropping, or when reference-image identity is the whole point.

Choose Wan 2.7 when you need to define the opening and closing frame, when you are cloning a specific voice, or when you want instruction-based editing of an existing clip. Frame control is the one thing HappyHorse does not offer as a first-class feature.

Use both when you are producing a sequence: Wan for the frame-locked product shots, HappyHorse for the atmosphere and dialogue shots, then cut them together. For a closer look at how it stacks up against other models, see Best Image to Video AI Tools in 2026. If you want to test it without installing anything, the free image-to-video generator works in the browser.

Free and cheap ways to try it

The cheapest test is not the API. It is a free front end with a short clip.

1. Generate a 3-second, 720p test on a free tool. Three seconds is enough to judge motion, audio timing and whether the model respects your framing. On a free Wan front end that costs nothing.
2. Run the same prompt at 1080p before you commit to a paid route. Resolution changes detail, not composition — but it also changes how long you wait.
3. Compare against Wan on the same brief. Use a free Wan video generator for the baseline, then judge whether HappyHorse's audio is worth the switch.

If you want the model itself rather than a proxy, per-second access is the sane entry point: you pay for what you generate, there is no monthly floor, and the retry cost is visible before you click.

Run Wan 2.7 Image to Video on Pollo →

Four mistakes that waste credits

1. Nine mediocre references. R2V does not average your images; it inherits their problems. A blurry reference gives you a confidently blurry character.

2. Writing a story instead of a shot. A 200-word narrative produces one shot trying to do everything. Write one shot per generation and stitch.

Want to see the difference on your own footage? Start creating with Wan 2.7 →

3. Ignoring the audio clause. If you do not describe the sound, you get generic ambience that has to be replaced — which wastes the model's main advantage.

4. Expecting a 30-second take. Fifteen seconds is the ceiling. Plan sequences in 10–15 second shots from the start rather than discovering the limit at the end.

The Bottom Line

HappyHorse 1.1 is the strongest argument yet that video generation is becoming a commodity input rather than a premium service. Native audio, nine references, 24fps at 1080p and a per-second price near Wan's — that combination used to cost a subscription and a queue.

What it does not do is replace a control-first workflow. No first/last frame, a 15-second ceiling, and dialogue-driven audio mean it complements Wan rather than displacing it.

The practical move is to stop choosing one. Run the audio-heavy shots through HappyHorse, run the frame-locked product shots through Wan, and keep both budgets per second so the comparison stays honest. Start with a free Wan test clip as your baseline, then spend on the model that actually differs on your brief.

FAQ

What is HappyHorse 1.1?

It is Alibaba's AI video generation model, an upgrade to HappyHorse 1.0. It generates video and synchronized audio in a single pass from text, an image, or up to nine reference images, at 720p or 1080p, in clips of 3 to 15 seconds.

How much does HappyHorse 1.1 cost?

Through per-second API access, about $0.1125 per second at 720p and $0.145 per second at 1080p. A 15-second 1080p clip is roughly $2.18 per attempt, before retries.

Is HappyHorse 1.1 better than Sora or Seedance?

It is positioned differently. HappyHorse generates audio and video together with strong reference-based identity at a low per-second price. Seedance 2.5 still wins on duration — 30 seconds in one pass — and heavy multimodal referencing. For most 5–15 second work, HappyHorse's cost advantage is the deciding factor.

Does HappyHorse 1.1 support 4K?

The documented output tiers are 720p and 1080p at 24fps. Some marketing pages reference higher-resolution ambitions for the model family, but plan for 1080p as the working ceiling.

Can I use HappyHorse 1.1 output commercially?

Commercial use is generally permitted on paid API tiers, and the 1.0 model family shipped with commercial usage rights for self-hosted deployments. Confirm the terms of the specific provider you use, since reseller licences vary.

What is the maximum clip length?

15 seconds per generation, as a whole number of seconds between 3 and 15. Longer videos require multiple generations stitched in an editor.

Does it work for talking-head videos?

Yes — native audio and multilingual lip-sync are the model's strongest differentiator, and reference images can hold a character's face across shots. If you need to clone a specific voice rather than generate a new one, Wan 2.7's voice cloning is the better fit.

References

• VentureBeat — Alibaba's AI video model rises to No. 2 in global rankings: 1.1 launch context, multi-reference identity, lip-sync improvements, 15B-parameter model family, Alibaba Cloud Model Studio distribution.
• Kie.ai — HappyHorse 1.1: per-second pricing at 720p and 1080p, aspect ratio list, duration range, reference image requirements.
• Artlist — HappyHorse 1.1 model page: 720p/1080p at 24fps, 3–15 second clips, up to nine reference images, prompting guidance.
• HappyHorse — model overview: 15B parameters, 40-layer unified transformer, DMD-2 distillation, seven-language lip-sync, open-source 1.0 release.

Related guides

• Happy Horse AI video generator guide — the 1.0 workflow this release builds on.
• Wan 2.7 vs HappyHorse 1.0 — how the previous generation compared on control.
• HappyHorse 1.0 — architecture and open-source details.
• Wan 3.0 complete guide — the closest price competitor at 1080p.
• Wan 3.0 vs Seedance 2.5 — where the long-clip premium is worth paying.
• Free AI video generator without watermark — free routes if you are still testing.

Generate a Free Test Clip Now →