Model reference · Updated July 2026
HappyHorse-1.0 Specs & Architecture
Everything published about how the model works, split into what is verified (observable via the hosted API) and what is reported (team disclosures awaiting the official repo).
Architecture (reported)
From team disclosures and technical press — unaudited until the official repo ships.
| Parameters | ~15 billion |
| Backbone | Single-stream transformer, 40 layers, unified self-attention (no dedicated cross-attention) |
| Layer layout | First 4 + last 4 layers use modality-specific projections; middle 32 share parameters across modalities |
| Modalities | Text, image, video, and audio tokens processed in one sequence — audio and video generated jointly in a single forward pass |
| Sampling | 8 steps via DMD-2 distillation; no classifier-free guidance required |
| Compiler | MagiCompiler full-graph optimization |
| Extras | Separate super-resolution module for upscaling to 1080p |
Performance (reported)
| 1080p generation | ~38 seconds for a 5-second clip on one H100 80GB |
| Draft speed | ~2 seconds for a 5-second clip at 256p |
| Native resolution | Up to 1080p, no upscaler required |
Capabilities (verified via hosted API)
| Modes | Text-to-video, image-to-video, video editing, reference-to-video |
| Clip length | 3–15 seconds |
| Aspect ratios | 16:9, 9:16, 1:1, 4:3, 3:4 |
| Audio | Joint audio-video generation with native lip-sync in 7 languages: Mandarin, Cantonese, English, Japanese, Korean, German, French |
| Styles | 50+ (cinematic, anime, documentary, fashion, …) |
Benchmarks
| Artificial Analysis Video Arena | #1 in both text-to-video and image-to-video (no-audio) at April 2026 debut — the only model to top both simultaneously |
| vs Seedance 2.0 | ~60 Elo ahead without audio; behind or tied with audio enabled |
Team & provenance
| Developer | Alibaba — Taotian Group's model lab |
| Lead | Zhang Di — 15-year AI veteran, ex-VP at Kuaishou, technical architect of Kling AI; rejoined Alibaba in late 2025 |
| Public debut | April 7, 2026 (anonymous arena entry) |
| Attribution confirmed | April 10, 2026 |
Planning to run it yourself? See the self-hosting guide for hardware requirements, or check what is verifiably released today.