On April 8, a horse suddenly appeared on the Artificial Analysis video arena leaderboard.
An anonymous model called HappyHorse scored Elo 1333 for text-to-video and 1391 for image-to-video, both breaking records and knocking ByteDance's SEEDANCE 2.0 off the top spot. SEEDANCE 2.0 had just launched in March 2026, beating out Google Veo 3, OpenAI Sora 2, and Runway Gen-4.5 to claim first place. Then along came a little pony and flipped the table.
The Tradition of Anonymous Benchmarking
This kind of anonymous leaderboard drop has become something of a recurring act in China's AI circle.
In February, an anonymous language model called Pony Alpha appeared on OpenRouter, free to use with a 200K context window, processing 40 billion tokens on its first day. Five days later, Zhipu AI announced that Pony Alpha was actually GLM-5, a 745B MoE architecture. In March came Hunter Alpha, which the community initially guessed was DeepSeek V4, until Xiaomi stepped forward to claim it—MiMo-V2-Pro, sporting one trillion parameters.
The benefits of anonymous benchmarking are straightforward: you get real blind-test data without the halo or baggage of a brand name. Benchmark scores can be gamed; blind tests can't.
HappyHorse followed the same playbook. But one detail quickly gave it away—Chinese and Cantonese were listed first among supported languages.
The Domain Name Rush
Once the model went viral, domain names were naturally targeted.
When I searched for HappyHorse's official website, I found something hilarious: both happyhorse.io and happyhorse.com had already been squatted, with full websites set up and charging users right away. Clicking in, they offered the whole package—text-to-image, text-to-music, text-to-video—putting on quite a show. But a closer look revealed they weren't using the HappyHorse model at all; the backend was actually running Lightricks' LTX—an open-source model from an Israeli company, originally only 2B parameters. I had tested it before, and it was completely different from the HappyHorse that topped the leaderboard.
Domain squatting is much faster than model training. But if someone unaware pays money thinking they're using the number-one HappyHorse from the leaderboard, that's a bit of a scam.
It wasn't just domains. Several HappyHorse-related repositories also popped up on HuggingFace—happyhorse-lab, happyhorseai, HappyHorseOrg—looking very official. But checking their creation dates, they were all registered on April 9. Clicking in, they either had only a README or were empty repos. The READMEs were quite complete, with words like "open source" and "number one," but no weight files. The hype-chasers weren't just squatting domains anymore; they were occupying HuggingFace too.
The Mystery Remains Unsolved
As of writing, HappyHorse's origin still has no definitive answer.
The model page on Artificial Analysis still says "More details coming soon," using the placeholder image for a mystery model. The leaderboard recognized its scores but provided no team background. There is also no official technical report, GitHub repository, company announcement, or paper landing page to close the identity loop.
The most plausible inference right now points to Sand.ai. The technical descriptions circulating for HappyHorse—15B parameters, 40-layer single-stream Transformer, joint text-video-audio modeling, 8-step DMD-2 distillation, multilingual lip-sync—heavily overlap with daVinci-MagiHuman, jointly released by Sand.ai and SII-GAIR. A 36Kr report also points in this direction. But so far, this remains speculation, not official confirmation.
The so-called "already open-sourced" claim also deserves a question mark. Artificial Analysis marks models with open weights as Open Weights on the leaderboard; HappyHorse currently lacks this label. The leading open-source video models are still around the LTX-2 Pro tier. Online articles claiming HappyHorse has been fully open-sourced under Apache 2.0 currently don't match any verifiable weight release.
Around the same time, Alibaba released Wanxiang 2.7, a 27B-parameter MoE architecture (14B active), supporting a "thinking mode." But Wanxiang 2.7 currently only offers an API; the weights have not been released. Previous Wanxiang models were open-sourced upon release; this time, for some reason, they weren't.
The Hard Need for Private Text-to-Video Deployment
HappyHorse's identity will be revealed sooner or later. But what interests me more isn't who built it, but the logic behind privatizing text-to-video models.
Every model type ignites a corresponding hardware category. When DeepSeek dropped, H20 orders exploded—in Q1 2025 alone, Chinese firms placed over $16 billion in orders. After open-source language models took off, DeepSeek V3 was shown running on a cluster of eight M4 Pro Mac Minis, and Mac Minis promptly sold out.
What will text-to-video ignite? I believe the answer is consumer-grade GPUs and small inference boxes. Moreover, the hard need for private text-to-video deployment is much stronger than for language models.
Delay-Tolerant, Cost-Sensitive
Text-to-video is naturally a "can-wait" scenario. Generating a video in the cloud already takes several minutes. Running it locally a bit slower—ten or even thirty minutes—makes no fundamental difference. You won't stare at the progress bar; you'll go do something else.
If latency isn't sensitive, then cost is. Besides compute, there's another major cost of cloud video generation that's easy to overlook: bandwidth. Videos are easily tens or hundreds of megabytes; moving them back and forth incurs frightening network fees. I recently did the math—the server itself isn't that expensive, but the bandwidth bill is something you only want to look at once. Keep it local, and that money stays in your pocket.
Delay tolerance also leads to another conclusion: you don't need top-tier compute. Language model inference chases low latency, demanding the best cards. Text-to-video is different; a bit slower is fine. Compute that is "not fast enough but cheap enough"—gaming GPUs, previous-gen compute cards—becomes a high-ROI choice.
The Dilemma of Regulation and Content Moderation
This is something many people haven't thought much about.
Text content moderation is manageable; most scenarios don't run into legal issues. Images and videos are different. IP infringement, portrait rights, sensitive content—the regulations are still evolving, and cloud service providers are stuck in a tough spot.
The dilemma for cloud providers is: if they don't block, they bear liability when something goes wrong; if they do block, they can't filter with enough precision technically, so they end up with massive false positives. The result is a "better to kill a thousand than miss one" approach—no uploaded portraits, no specific IPs allowed, freeze at the slightest hint of sensitivity. The user experience becomes full of restrictions.
Local deployment avoids these problems entirely. The model runs on your own machine, bypassing third-party review. In the Stable Diffusion era, a huge amount of text-to-image workflows ran locally—not because local was faster, but because there were no moderation restrictions. Text-to-video will repeat this pattern.
Shorter Path to Monetization
The value of language models has always been hard to quantify. A better model writes a better paragraph—how much revenue does that bring? It's unclear. Upgrading from a 32B model to a hundreds-of-billions-parameter private deployment, spending ten times more on H20s—does that earn ten times more? No one can say for sure. The coding scenario improved things somewhat, but before that, people genuinely couldn't make the math work.
Text-to-video is completely different. A good video is good traffic, and traffic is money. Spend a few hundred to generate a decent-quality video; if the content is interesting, the traffic it brings could be worth thousands or even tens of thousands. Anyone can do that math.
SEEDANCE 2.0 is a case in point. Creators are willing to pay and queue for resources because videos produced with it genuinely drive better metrics. The gap between good and bad models becomes obvious after just a few posts.
Hardware Ripple Effects
Whether HappyHorse will actually open-source, and when, is still uncertain. But we can already do some math.
If we estimate based on the rumored 15B parameters, FP16 inference would require roughly 30 GB of VRAM, while INT8 quantization needs only around 15 GB. A single RTX 4090 or 5090 could handle it. Something like the DGX Spark, a small box with 128 GB of unified memory, would be more than comfortable running inference.
If it were actually open-sourced at this size, RTX 4090/5090 cards would likely become even harder to buy. The DGX Spark has already risen from its original 2025 announcement price of 4,699](https://www.tomshardware.com/desktops/mini-pcs/nvidia-dgx-spark-gets-18-percent-price-increase-as-memory-shortages-bite-founders-edition-now-usd4-699-up-from-usd3-999)—an increase of over 50%—and supply was already tight. Adding another memory-hungry heavyweight to the mix will only make things more extreme.
We've seen this script before. DeepSeek ignited H20 demand; open-source LLMs drove Mac Mini sales. Text-to-video has reached this level of quality; all that's missing is a good enough open-source model to land. Whether HappyHorse gets that opportunity remains to be seen, but it will happen sooner or later.
Still Up in the Air
Back to HappyHorse itself.
Will it officially open-source? There's no telling right now. The leaderboard scores are there, but weights and code haven't materialized. If it ends up being just an API service, the impact on the hardware market will be limited—just another powerful closed-source model.
How big is it really? Marketing pages say 15B; if true, a single consumer GPU can run it. But if it's actually larger, requiring multi-GPU or even a cluster, then local deployment becomes unrealistic, and we're back to the cloud provider model.
Different answers to these two questions lead to completely different storylines. But regardless of how HappyHorse turns out, the trend of moving text-to-video local won't change. Tools like ComfyUI and WebUI are waiting for a good enough open-source model, and so is the quantization community. Once it arrives, the consumer hardware side is going to get lively.