Is HappyHorse-1.0 actually open source? What can be verified
Alibaba announced an open-source release — base model, distilled variant, super-resolution module, Apache 2.0. Here is what independent checks actually found, kept up to date.
Short answer, as of July 2026: announced, not independently verified. If your plan depends on self-hosting HappyHorse-1.0, don't commit until the checklist below turns green.
What was announced
Alongside the April 10, 2026 reveal, Alibaba announced open-source plans for HappyHorse-1.0 covering:
- the ~15B base model weights
- the DMD-2 distilled fast variant
- a super-resolution module
- inference code
- an Apache 2.0 license permitting commercial use
Hugging Face organization pages for the model exist and list these artifacts.
What independent checks found
Multiple third-party investigations in April and May 2026 reported the same pattern:
- Weight files returned access errors (HTTP 401) when download was attempted, rather than downloading.
- The GitHub repository was readable but empty of substance — no weights, no inference code, and no license file at the time of checking.
- No independently confirmed local deployment — no reproducible community writeup of running the model on self-managed hardware has surfaced.
None of this proves bad faith — staged releases, legal review, and safety gating routinely delay published weights. But "open source" today describes an announcement, not an artifact you can download.
The verification checklist
The claim becomes verified when all four hold. Status as of 2026-07-17:
| Check | Status |
|---|---|
| Weights downloadable without special access | ❌ Not confirmed |
| Inference code published and runnable | ❌ Not confirmed |
| License file present in the released repo | ❌ Not confirmed |
| Independent community reproduction of local inference | ❌ Not confirmed |
We revisit this page when the status changes; the updated date above reflects the last check.
What to do meanwhile
- Build on hosted APIs — the official Bailian API, fal.ai, or aggregators. Our API lesson covers endpoints and per-clip cost math.
- Treat "self-hostable" claims in vendor marketing with suspicion — as of now, nobody outside Alibaba has demonstrably self-hosted this model.
- If weights do land, expect the distilled DMD-2 variant (8-step generation) to be the practical one for single-GPU use — the team's own numbers cite ~38 s per 1080p clip on an H100.
For the full background on the model itself, start with Lesson 1: What is HappyHorse-1.0?