Happy-Horse-1.0: a README collecting claims about an unreleased video model

One file, and a disclaimer that defines the whole project

The repository contains exactly one entry, a README. There is no code, no model weights, no inference script and no configuration, and the project description says as much: information collection for the Happy Horse AI video generator model, with the official demo and updates at happyhorses.io.

The disclaimer at the top is unusually blunt and it governs everything else in the file. This is a personal information-collection repository, not an official Happy Horse repository and not a public release. Happy Horse 1.0 has not yet been officially open-sourced, and no model weights, no inference code and no official GitHub repository have been published. The author states they are not affiliated with the Happy Horse team, that the information may be incomplete, inconsistent across sources, or change at any time, and that they will sync the README once the official model, repository or technical report is released.

The sources are listed too, and they are a mixed bag: community-compiled architecture notes, alleged technical leaks, the Artificial Analysis Video Arena, project landing pages and independent discussion threads. The collection is for personal reference and learning.

The practical reading of this is that the repository is a status tracker, not a project. It is worth reading if you want to know what has been said and how confident the various claims are, and it is worth ignoring if you were looking for something you can install. The last push was on 2026-05-12 and the file carries a date marker of 2026.4.27, and there are no GitHub releases, so the history is the commit log.

What the model is claimed to do, and what it costs

The README records that Happy Horse 1.0 is Alibaba's video generation and editing model, and that it started grayscale testing in China on April 27, 2026.

The listed capabilities are text-to-video generation, image-to-video generation, video editing, generation of 3 to 15 second videos, multi-shot generation, coherent short-form storytelling, resolution up to 1080p, and up to 4 videos generated at the same time. The strengths section lists fast generation at around 2 to 5 minutes per video, strong prompt following, good control over camera movement, composition, style and atmosphere, high reference fidelity in image-to-video, and support for multi-shot storytelling. The limitations are equally specific: audio-video synchronization still needs improvement in complex scenes, longer videos may contain physical inconsistencies, and text rendering may produce garbled or incorrect characters.

Pricing is given per second, in two columns, a listed price and a discounted Pro price. At 720p that is RMB 0.9 per second and RMB 0.44 per second, and at 1080p it is RMB 1.6 per second and RMB 0.78 per second. Those are the kind of numbers worth checking against the official platform before you budget anything, since the whole table comes from the same unverified collection.

Availability is listed through Alibaba Cloud Bailian, the official website, the Qwen App, Alibaba Wukong, MuleRun and JVS Claw, which is a broad set of distribution channels for a model in grayscale testing.

On the Artificial Analysis Video Arena, the README records a second place finish in text-to-video, image-to-video and video editing, behind ByteDance Seedance 2.0.

The architecture table is explicitly unverified, and says so

The technical section is the most interesting part of the file and the part the disclaimer is aimed at.

The table is introduced with a warning that the details come from community-compiled architecture notes and alleged technical leaks circulating about the model, that none of it has been confirmed by an official technical report or peer review, and that it should be treated as plausible but unverified until the model is released.

What it claims is specific enough to be checked later. Total parameters of about 15B. A unified self-attention Transformer, reportedly without dedicated cross-attention branches. Forty layers in total, laid out as a sandwich where the first 4 and last 4 layers use modality-specific projections and the middle 32 layers share parameters across all modalities. Text, image, video and audio tokens concatenated into a single token sequence. Multimodal fusion through per-attention-head learned scalar gates with sigmoid activation, described as being for training stability. Conditioning for the reference image and denoising signals routed through one minimal unified interface, with reportedly no dedicated conditioning branches. Timestep handling with reportedly no explicit timestep embeddings, the model said to infer the denoising state from the noise level of the input latents. DMD-2 distillation, Distribution Matching Distillation v2, at 8 sampling steps with reportedly no classifier-free guidance required. An inference runtime called MagiCompiler doing full-graph compilation with operator fusion across Transformer layers, reported to give roughly 1.2 times end-to-end speedup. And a reference GPU of NVIDIA H100 80GB.

The section then explains why the reported choices might matter, which is the most useful prose in the file. The unified self-attention claim is contrasted with the usual open-source video models, Wan 2.2, HunyuanVideo, LTX-2 and CogVideoX, which use a Diffusion Transformer backbone where text conditioning arrives through cross-attention from a separate text encoder, and where audio, when present, is produced by a different model entirely. The reported claim for Happy Horse is that audio-video alignment is learned as part of denoising rather than bolted on afterwards. The sandwich layout is described as a parameter efficiency story, with most of the network doing cross-modal reasoning rather than being split into siloed sub-networks. The per-head gating is explained as a response to the instability of joint multimodal training.

A mystery model on a leaderboard, and a release promise

The origin story in the README is worth separating from the architecture claims, because it comes from a different kind of source.

Happy Horse is said to have first surfaced publicly as a mystery model on the Artificial Analysis Video Arena, where it appeared anonymously alongside frontier closed models from ByteDance, Kling and Google. Independent community observers noted that it appeared to come from Asia and that, unusually for a mystery entry, it included native audio output.

That native audio point is the substantive technical claim, and the README explains what it means. Happy Horse is described in the file as an open-source AI video generator with native joint audio-video generation, meaning the model produces video frames and the corresponding audio track, dialogue, ambient sound and Foley, together in a single forward pass, rather than generating silent video and dubbing it afterwards.

The second layer of the reported performance claim is the distillation result: combined with DMD-2, the distilled variant is reported to generate 1080p video in roughly 38 seconds on an NVIDIA H100, using only 8 denoising steps without classifier-free guidance. That figure is the one most likely to be quoted and the one that should be treated most carefully, because it depends on a runtime, MagiCompiler, and a hardware reference point, neither of which you can reproduce from this repository.

The announcement promise is listed in the same file: the public release will include the base model, the distilled 8-step model, the super-resolution module and the inference code. None of those have been published yet, and the file points to the official platform for release news.

How to use this repository, which is to read it and wait

There is no installation, and the honest way to describe using this project is to read it and come back later.

The official links given in the file are the model website at www.happyhorse.com and the platform page at happyhorses.io, and the project homepage is happyhorses.io. The community reference points are the Artificial Analysis Video Arena page for the video leaderboard and the independent discussion threads that fed the architecture notes.

The links listed under official websites, in the form they appear in the file, are:

- official website: [www.happyhorse.com](https://www.happyhorse.com/)
- Try HappyHorse now: [happyhorse](happyhorses.io)

What this repository is good for is being a single place where the claims are gathered with their confidence attached. Most coverage of a model like this is scattered across a leaderboard entry, a forum thread and a landing page, and this puts them in one file with a table of reported specifications and an explicit warning attached to it.

What it is not is a substitute for a technical report. The author says the same thing in different words, that the information may be incomplete, inconsistent across sources, or change at any time, and that they are watching for updates and will sync the README when the official model, repository or technical report is released. The licence is not identified and the primary language is not recorded, which is consistent with a repository that holds one prose file and nothing else.

Editorial conclusion

Happy-Horse-1.0 suits a reader who wants a single place to see what has been claimed about a model that has not launched, and the collection is unusually clear about its own status, flagging every technical claim as unverified and distinguishing what the announcement promises from what is actually published. It does not suit a developer looking for code to run, because the repository contains a README and nothing else, and it does not suit anyone who needs verified facts about the model, since the architecture table is built from community notes and alleged leaks. Before you rely on any of it, check the official platform for a technical report, and treat the parameter count, the layer layout and the inference timings as claims to be confirmed rather than specifications you can plan against.

Frequently asked questions

Official sources

1. brooks376/Happy-Horse-1.0 on GitHub
2. Issues
3. Project website
4. README