Is the dark horse that surpasses Seedance a "carefully designed" hit?

A folded chat record has once again set off a storm in the AI circle.

The cause is that netizens discovered a new video generation model called HappyHorse-1.0. It is not only powerful but also very mysterious.

It is powerful because it can surpass Seedance2.0 on the AA Ranking List (officially known as artificialanalysis.ai) and become the world's number one.

What's mysterious is that currently, no one has claimed this model, and there are no any identifiers related to its identity on its official website.

Many people, upon seeing this model, couldn't help but ask like Cao Cao: "I originally thought Lü Bu was invincible in the world, but unexpectedly, there's someone even more valiant! Whose general is this?"

Therefore, speculating about the ownership of HappyHorse-1.0 has become the hottest topic in the industry today.

Of course, "Youjie UnKnown" also actively rushed to the frontline of the gossip. However, during the process of following the gossip, we found that not only is HappyHorse outstanding in terms of capabilities, but the fact that it has become a big hit also shows signs of human manipulation.

Everything might be a well - designed scheme.

What makes HappyHorse so remarkable?

As a dark horse that suddenly emerged, HappyHorse 1.0 naturally has its remarkable features. Let's start with this.

First of all, different from mainstream video models, HappyHorse is open - source.

Secondly, it only has 15B parameters, which is smaller than most models, and it can complete ultra - fast inference in 8 steps (it only takes 2 seconds for 256p, and only about 38 seconds for 1080p). It was officially launched at dawn on April 8th.

Its main advantages mainly lie in two points: fast generation speed and excellent audio capabilities.

First, the generation speed is very fast.

In existing cases, with just one prompt and one picture, a complete video can be generated in about one minute. In the circulated WeChat chat records, there are also some remarks deliberately criticizing Jimeng.

Of course, the fast generation speed might be because the number of users is still small, so there's no need to wait in line.

Secondly, in terms of audio capabilities, HappyHorse also has more obvious improvements compared to most video models.

One is the matching of environmental sounds.

In the cases of HappyHorse, when a foot steps on the ice, you can hear a "bang"; when a basketball goes into the hoop, there will be a sound of collision with the hoop; when a barista pours milk, the sound of pouring milk can also be generated.

Previously, many AI - generated videos often had problems of out - of - sync and mismatched audio and video.

More importantly, the sounds in HappyHorse are not simply superimposed background sounds, but have a basically consistent corresponding relationship with the actions in the picture. In other words, the sound has become "a part of the video" rather than an added - later effect.

The second is HappyHorse's language ability.

After inputting a prompt, the model can directly generate voice content in the corresponding language. According to the official website information, currently, the main supported languages are Mandarin, Cantonese, English, Japanese, Korean, German, and French.

Finally, in terms of price.

HappyHorse uses a point - based system and offers three payment methods: monthly, annually, and on - demand.

Although people on the Internet are saying that its price is half that of Seedance 2.0, in fact, it is almost set according to the subscription price of Seedance 2.0 and is also higher than the pricing of Pixverse V6, so it doesn't have a price advantage.

All these have made HappyHorse a hot topic of discussion.

An artificially created "dark horse"?

Although most of the discussions about HappyHorse occurred naturally, various signs indicate that from deliberately hiding information to create topics for discussion, to releasing the folded chat record to expose and trigger discussions, this could well be a well - designed and artificial promotion.

First, regarding the ranking list, a friend working in an AI video company said that the AA ranking list can be manipulated. As long as there is a budget, they can provide the questions in advance. On this basis, if the samples are optimized accordingly, it's easy to get a high score.

HappyHorse quietly topped the list and then quickly dropped off. Now, all that's circulating in the market are screenshots, which also provides a basis for this speculation.

Especially amidst the almost one - sided praise, some netizens on X also said that from the generated effects, the sense of realism is relatively weak, so it's possible that Happyhorse's ranking was artificially inflated.

For example, some X users found that although the demos on the official website look good, there are some gaps in terms of physical performance. For instance, when presenting fast - moving objects, there are still AI motion artifacts, especially for objects with stripes or lines.

Secondly, HappyHorse became popular after topping the AA ranking list in just two days, and its official website already has a complete display of why it's popular, which is really efficient.

Third, the materials circulating on the entire network are very single.

"Youjie Unknown" tried to pay to test the video, but after multiple attempts, it couldn't be opened. The free points provided were not enough to support the test.

And the videos circulating online are also very single, basically the same as the content spread through the folded chat record in the WeChat group, without new and more video content.

This is quite interesting. When the whole network is saying that it's better than other models, the only evidence is the video cases provided by the official, which is really thought - provoking.

Interestingly, in this chat record, it implicitly and explicitly compares with competitors, highlighting the contrast and creating topics. I wonder how the competitors are feeling now.

Of course, all these flaws don't overshadow its merits.

Even if this is a well - planned marketing campaign, as long as the product itself is good, everything else is not a problem.

What does HappyHorse bring?

In fact, HappyHorse really brings something new, a new technological paradigm.

Currently, mainstream video generation models basically use the method of first compressing the video and then handing it over to the Transformer for "denoising and generation" step by step.

Different companies also have different specific operations.

The more traditional and currently mainstream method is to process the video and audio separately first and then integrate them. For example, HunyuanVideo, PixVerse, and the early Seedance 1.0 all do this. They first process the text, video, or space and time separately and then combine them later.

The unconventional method is to prioritize speed improvement. For example, LTX developed by the Israeli company Lightricks focuses not on the complexity of the structure but on compressing the video more severely to reduce the number of tokens, so that the generation is faster.

In the future, the evolving trend in the industry is: unified generation with a single large backbone.

In the past, the method of processing separately first and then integrating is usually called "two - stream" in technology.

The picture and sound are generated by different modules respectively and then pieced together through an alignment mechanism. This is why, in many cases, although the audio and video can be matched, there is still a bit of unnaturalness - because they were not generated together from the beginning.

Now, unified generation with a single large backbone tends to be a "single - stream" generation method, that is, the picture and sound are treated as different parts of the same generation process and completed together.

That is to say, the sound is not added later but "grows" together with the picture.

This difference can be seen from the specific effects. For example, the cracking sound when stepping on the ice and the collision sound when the basketball goes into the hoop are not simply superimposed but appear synchronously with the actions, showing an obvious causal relationship.

The latest models on the market, such as Sora, Veo, and Seedance 2.0, are all moving in this direction, which is also the current technological direction.

HappyHorse adopts this direction, but it goes even further, emphasizing "unified processing of everything together." (At least that's what they claim on their official website)

This is specifically manifested in:

1. More unified in structure, emphasizing single - stream and no cross - attention;

2. More unified in modality, not only processing text and video together but also trying to incorporate audio;

3. More radical in inference, generating in 8 steps, emphasizing low latency and fast speed.

To put it more simply, the efficiency of video generation is higher. For example, the queuing situation like in Seedance might be reduced. The consistency is better, and problems such as out - of - sync audio and video, out - of - sync lip - syncing, and incorrect action timing can all be improved.

Whose general is HappyHorse?

Happy Horse became well - known partly because of its capabilities and partly because of its mystery.

There is almost no effective information about its "identity" online. The introduction on its official website is also very reserved, only leaving a sentence: "Happy Horse 1.0 is developed by the Happy Horse team."

The more mysterious HappyHorse is, the more curious people become. So everyone is asking around about where HappyHorse comes from, and there are soon many speculations in the market.

Currently, the most widely believed one is that it comes from the Future Life Laboratory of Taotian Group led by Zhang Di. Zhang Di was previously in charge of the Kling project at Kuaishou, and before that, was the person in charge of the big data and machine - learning engineering architecture at Alibaba Mama.

There is a lot of information about this source, and some self - media have also reported on it, but the official has neither confirmed it nor is there an authoritative source of information.

Besides this, the first one to be suspected is the Kuaishou camp.

Some people noticed that there are traces related to spaceship.com left on the official website, and "spaceship" happens to coincide with the trademark of Kuaishou's affiliated company, so it is speculated that this project might come from Kuaishou.

The second one to be suspected is Tencent or Grok (because they are all related to horses).

In fact, we previously suspected it was Aishi, but after communication, they denied