Fact-checked against Alibaba's own documentation

HappyHorse AI: What Alibaba Actually Documents

Seven models across two generations — 1.1 and 1.0 — with 3 to 15 second clips at 720P or 1080P. It topped the blind-test rankings in April 2026; it does not today.

Most write-ups about HappyHorse quote a parameter count and a benchmark position that Alibaba has never published, and several still say it has no API. This page separates what QwenCloud's documentation states from what is being repeated, names the source for every line, and dates every check. Last verified August 7, 2026.

HappyHorse is not in our generator — but it is publicly available

You cannot generate with HappyHorse on ImagineToVideo. You can use it elsewhere: fal has offered four HappyHorse endpoints since April 27, 2026, and Alibaba Cloud Model Studio (Bailian) offers it too. Both need an API key and a developer setup. This page documents what Alibaba's own docs say; it is not a pitch to avoid the model.

All sources last checked on

HappyHorse is an API. Some jobs want a browser.

If you would rather write a prompt and get a clip back without keys or SDKs, Seedance 2.5, Kling V3, VEO 3.1 and Seedance 2.0 run here.

Open the generator

What QwenCloud's documentation states

3 to 15 seconds per clip

Every documented HappyHorse model — both generations, all four modes — carries the same 3–15 second range. There is no longer tier in the documentation.

720P or 1080P

The documented resolution options are 720P and 1080P for all seven models. No 4K tier appears anywhere in the documentation.

Audio on every mode

Text-to-video, image-to-video and reference-to-video all list audio. The video-edit model can either generate new audio or keep the original track.

One to nine reference characters

The reference-to-video models accept 1–9 references for multi-character work — the widest character-reference range Alibaba documents. Video-edit takes 0–5 reference images.

Confirmed, with the source for each line

Every statement below comes from Alibaba's own documentation or from the named news report. Nothing here is sourced from an aggregator or a spec-summary page.

  • Alibaba documents seven HappyHorse models across two generations.

    happyhorse-1.1-t2v, 1.1-i2v and 1.1-r2v, plus 1.0-t2v, 1.0-i2v, 1.0-r2v and 1.0-video-edit. Most coverage still describes HappyHorse as a single 1.0 model.

    QwenCloud video models documentation ·

  • Clip length is 3 to 15 seconds across every documented model.

    The range does not change between 1.0 and 1.1, nor between text-to-video, image-to-video, reference-to-video and video-edit.

    QwenCloud video models documentation ·

  • The documented resolutions are 720P and 1080P.

    This applies to all seven models. No 4K option is documented for any of them.

    QwenCloud video models documentation ·

  • Reference-to-video accepts 1–9 references; video-edit accepts 0–5 reference images.

    The r2v entries are documented for multi-character work. The video-edit entry additionally documents an audio choice between generated and original.

    QwenCloud video models documentation ·

  • The documentation carries no invite-only flag and no pricing.

    Unlike some other models on the same platform, HappyHorse is not marked as gated. No rate appears in the documentation either.

    QwenCloud video models documentation ·

  • fal has offered four HappyHorse endpoints since April 27, 2026.

    Text-to-video, image-to-video, reference-to-video and video-edit, launched as an official API partnership. Alibaba Cloud Model Studio (Bailian) offers the model as well.

    fal HappyHorse API ·

  • Alibaba confirmed ownership on April 10, 2026, after an anonymous debut.

    The model appeared on the Artificial Analysis Video Arena around April 7 without naming its creator and reached the top of the blind-test rankings before Bloomberg and CNBC reported Alibaba's confirmation.

    CNBC, April 10, 2026 ·

All seven documented models

Straight from QwenCloud's documentation. We have not found another page that lists the full line-up, which is why the exact model identifiers matter.

happyhorse-1.1-t2v
Mode
Text to video
Length
3–15s
Resolution
720P, 1080P
Documented features
Audio, aspect ratio control
happyhorse-1.1-i2v
Mode
Image to video
Length
3–15s
Resolution
720P, 1080P
Documented features
Audio, first-frame input
happyhorse-1.1-r2v
Mode
Reference to video
Length
3–15s
Resolution
720P, 1080P
Documented features
Audio, 1–9 character references, aspect ratio control
happyhorse-1.0-t2v
Mode
Text to video
Length
3–15s
Resolution
720P, 1080P
Documented features
Audio, aspect ratio control
happyhorse-1.0-i2v
Mode
Image to video
Length
3–15s
Resolution
720P, 1080P
Documented features
Audio, first-frame input
happyhorse-1.0-r2v
Mode
Reference to video
Length
3–15s
Resolution
720P, 1080P
Documented features
Audio, 1–9 character references, aspect ratio control
happyhorse-1.0-video-edit
Mode
Video editing
Length
3–15s
Resolution
720P, 1080P
Documented features
Audio generated or original, 0–5 reference images

Line-up as documented on 2026-08-07. Re-checked weekly; a new generation would appear here first.

Claims in circulation, checked one by one

These are the HappyHorse specs being repeated across the web. For each one, what Alibaba's documentation actually says.

Claim in circulationWhat the documentation saysVerdict
15 billion parameters, 40-layer self-attention Transformer
⚠️ Unconfirmed — no official figure exists
Roughly 38 seconds for 1080p on a single H100
⚠️ Unconfirmed — no official benchmark published
Audio and video generated jointly in a single forward pass
⚠️ Unconfirmed as an architecture claim

That audio exists is confirmed. How it is produced is not.

There is only a 1.0 version
Contradicted by the documented line-up
It has no public API and cannot be used
Contradicted — it is publicly available

An earlier version of our own HappyHorse article said this. It was corrected on 2026-08-07.

Open source, with weights already released
⚠️ Unconfirmed — nothing published so far

Every verdict reflects the state of the sources on 2026-08-07 and is re-checked weekly.

The "number one" claim, with a timestamp

HappyHorse really did top the rankings. That was four months ago. Both halves of that sentence matter.

#1 on Artificial Analysis text-to-video
Was it right then?
Yes — accurate at the time
As of 2026-08-07
#5 HappyHorse-1.1 (Elo 1148) and #6 HappyHorse-1.0 (1127). The leader is Gemini Omni Flash at 1244.
Ahead of Seedance 2.0 in blind tests
Who said it, and when
Widely reported from the April 2026 arena data
Was it right then?
Yes — accurate at the time
As of 2026-08-07
Reversed. Dreamina Seedance 2.0 720p is #3 at Elo 1224, ahead of both HappyHorse entries.

Elo is a relative score: everyone's number moves when new models enter the arena, so a drop in rank does not by itself mean a model got worse. Check the live leaderboard rather than any figure quoted in an article — including this one.

What this means for you

If you write code and need multi-character control

HappyHorse's reference-to-video models take 1–9 character references, and the video-edit model takes 0–5 reference images. That is the most explicit multi-character surface Alibaba documents. If that is your job and an API key is not an obstacle, go use it on fal — we would rather tell you that than pretend it does not exist.

If you want a clip without building anything

HappyHorse ships as an API on both platforms that carry it: keys, per-call billing, no interface. If you want to type a prompt in a browser and get a clip back, that is a difference in product shape, not in model quality — and it is the gap the models on this site fill.

How to check any HappyHorse claim yourself

Four steps, no special access. If a spec cannot survive these, it is not a spec.

1

Open the model table

Go to QwenCloud's video models documentation and find the HappyHorse rows. Length, resolution and per-mode features are all there — and so is the version line-up, which is how you catch articles still describing 1.0 as the only release.

2

Check whether the API is still live

Open fal's HappyHorse endpoints. Availability changes; an article claiming the model cannot be used is only ever true on the day it was written.

3

Look up today's ranking

Open the Artificial Analysis text-to-video leaderboard and read the current position. Any "#1" without a date attached is telling you about the past.

4

Trace every number to Alibaba

For parameter counts, layer counts and generation timings, follow the claim back until you reach Alibaba's own documentation. For this model those numbers stop at somebody else's summary page.

Update log

The three sources are re-checked weekly. Every change lands here with its date.

  1. Page published; two widely repeated claims corrected

    The documentation was found to list 1.1 as well as 1.0, which contradicts the common description of a single release, and fal's endpoints were confirmed live, which contradicts the "no public API" claim — including in an earlier version of our own article, corrected the same day. Current ranking recorded as #5 and #6.

    QwenCloud video models documentation

  2. fal launches HappyHorse-1.0 with four endpoints

    Text-to-video, image-to-video, reference-to-video and video-edit, as an official API partnership. The model is also available through Alibaba Cloud Model Studio.

    fal HappyHorse API

  3. Alibaba confirmed as the creator

    Bloomberg and CNBC reported Alibaba's confirmation after the model had reached the top of the blind-test rankings under no name.

    CNBC

Questions people are asking about HappyHorse

What is HappyHorse AI?

A family of video generation models from Alibaba. QwenCloud's documentation lists seven: happyhorse-1.1-t2v, 1.1-i2v and 1.1-r2v, plus 1.0-t2v, 1.0-i2v, 1.0-r2v and 1.0-video-edit. All are documented at 3–15 seconds, 720P or 1080P, with audio.

How many versions are there?

Two generations. Most coverage describes HappyHorse as a single 1.0 model, but the documentation lists 1.1 in three modes — text-to-video, image-to-video and reference-to-video — alongside the four 1.0 models. Both generations appear on the current leaderboard.

Can I use HappyHorse right now?

Yes, through an API. fal has offered four endpoints since April 27, 2026, and Alibaba Cloud Model Studio carries the model too. Both need an API key and a developer setup. You cannot use HappyHorse on ImagineToVideo — it is not in our generator.

Does HappyHorse support 4K?

Not according to the documentation. Every one of the seven models is documented at 720P or 1080P, and no 4K option appears anywhere. If you need 4K today, VEO 3.1 offers it on this site.

Is HappyHorse 15 billion parameters?

⚠️ Unconfirmed. That figure is widely repeated but does not appear in Alibaba's documentation, and Alibaba has not published a technical report for this model. The same applies to the 40-layer architecture and the roughly-38-seconds-on-an-H100 timing.

Is HappyHorse open source?

⚠️ Not confirmed. The documentation covers the API only — it does not mention weights, a licence, or an open-source plan, and no weights have been published as of August 7, 2026. We re-check weekly.

Is HappyHorse still the number one model?

No, and it genuinely was. It topped the Artificial Analysis blind-test rankings around April 2026. As of August 7, 2026 it sits at #5 (HappyHorse-1.1, Elo 1148) and #6 (1.0, Elo 1127), with Gemini Omni Flash leading at 1244. Elo shifts as new models enter, so a lower rank is not by itself evidence that a model got worse.

Will ImagineToVideo add HappyHorse?

We will evaluate it based on real demand. If enough people tell us they want HappyHorse specifically — particularly its multi-character reference mode — we will look at adding it; whether we can depends on official pricing and terms, which Alibaba has not published. In the meantime the models on this site are the ones we can stand behind today.

HappyHorse is an API. This is a browser.

If you need a clip today without keys, SDKs or per-call billing, Seedance 2.5, Kling V3, VEO 3.1 and Seedance 2.0 all run here — with their real parameter limits listed.

HappyHorse AI: All 7 Models, Official Specs Checked