HappyHorse-1.0

How Artificial Analysis Works

Artificial Analysis runs a blind comparison arena. Users are shown two video outputs generated from the same prompt — they don't know which model produced which output. They simply vote on which video looks better. These votes are then fed into an Elo rating system, the same mathematical framework used in chess rankings.

This methodology matters because it eliminates self-reported benchmarks. When a company says their model scores 95/100 on their internal test suite, that number is marketing. When thousands of anonymous users independently prefer one model's output over another in blind tests, that's a market signal.

What HappyHorse's Elo Scores Actually Tell Us

An Elo difference of ~60 points translates to roughly a 58-59% win rate — meaning in a head-to-head comparison, the higher-ranked model would be preferred about 6 out of 10 times. HappyHorse's lead over the #2 model (Seedance 2.0) is 97 points in T2V — a statistically substantial gap.

However, Elo scores for newly added models are inherently volatile. Seedance 2.0 has accumulated over 7,500 vote samples, establishing a stable rating. HappyHorse's sample count is still growing. As more votes come in, the score could stabilize higher, lower, or roughly where it is.

Single Self-Attention Transformer (Claimed)

According to information on happyhorses.io, HappyHorse-1.0 uses a unified 40-layer Transformer architecture. Unlike traditional multi-modal models that use separate encoders with cross-attention bridges, HappyHorse reportedly processes all modalities — text tokens, reference image latents, and noisy video/audio tokens — through a single self-attention mechanism.

The first and last 4 layers allegedly use modality-specific projections (mapping each data type into a shared embedding space), while the middle 32 layers share parameters across all modalities. This design, if true, would be architecturally elegant — it means the model learns unified representations rather than bolting together separate subsystems.

How This Compares to Known Architectures

If the claims are accurate, HappyHorse's architecture resembles an evolution of the single-stream approach seen in models like Alibaba's WAN series, but with joint audio-video denoising — a capability most competitors implement as a post-processing step.

For context: Seedance 2.0 uses a Dual-Branch Diffusion Transformer with an "Attention Bridge" connecting separate video and audio branches. Kling 3.0 uses a cascaded approach with separate super-resolution stages. HappyHorse's claimed single-stream design is arguably more ambitious, but also harder to verify without open weights.

Claimed Inference Performance

The primary HappyHorse site lists specific inference speeds: 2 seconds for a 5-second 256p clip, and 38 seconds for 1080p resolution on an H100 GPU. These numbers, if accurate, would make it one of the fastest high-quality video generators available.

However, these are self-reported vendor numbers with zero third-party verification. Until independent benchmarks are run on publicly available weights, these figures should be treated as marketing claims.

Text-to-Video Rankings (No Audio) — April 2026

Image-to-Video Rankings (No Audio) — April 2026

The Quality Signal Is Real

Regardless of who made HappyHorse, the Elo signal from blind voting is genuine. Thousands of users, without knowing the model's identity, consistently preferred its outputs. This isn't marketing — it's empirical preference data. Something capable has been built.

But You Can't Use It Today

For anyone building a pipeline, shipping a product, or creating content professionally: HappyHorse-1.0 doesn't exist as an option yet. No API, no weights, no playground, no pricing. The quality signal is interesting; the practical utility is zero.

What You Should Actually Do