Deployment planning · Updated July 2026
Self-Hosting HappyHorse-1.0
What your hardware needs to look like, which checkpoint to pick, and what you can prepare today — so you can deploy the day the weights are fully, verifiably out.
Before you rent a GPU: the open-source release is still in flux — repos have appeared under more than one Hugging Face org and the license file hasn't been verified. Check the open-source status tracker for what is actually downloadable today.
Hardware & software requirements
Numbers below are reported by the team and early testers — treat them as planning figures until the official repo ships benchmarks.
| Python | 3.10+ |
| CUDA | 11.8+ |
| Minimum VRAM | 24GB+ (distilled checkpoint, FP8 quantization) |
| Comfortable VRAM | 48–80GB (A100 / H100) for base-model 1080p |
| Reference speed | ~38s for a 5-second 1080p clip on one H100 80GB |
| Fast path | ~2s for a 5-second clip at 256p (draft previews) |
Which checkpoint should you run?
Base model (~15B)
Full quality, full VRAM appetite. The checkpoint the arena numbers come from. Expect A100/H100-class hardware for native 1080p.
Distilled 8-step variant
DMD-2 distillation cuts sampling to 8 steps with no classifier-free guidance. The realistic single-GPU option — pair it with FP8 quantization on a 24GB card.
Super-resolution module
Generate at a lower base resolution, then upscale to 1080p. Trades a little fidelity for a lot of VRAM headroom.
Self-hosting vs. the API: the break-even math
A rented H100 runs roughly $2–3/hour on major clouds. At ~38 seconds per 5-second 1080p clip, one H100-hour produces about 90–95 clips — call it ~$0.03 per clip in raw compute, versus $1.40 for the same clip on the hosted API.
The API still wins below a few hundred clips per month once you price in your time: deployment, queueing, retries, upgrades, and idle GPU hours are real costs. Past a few thousand clips per month — or if you need fine-tuning, custom pipelines, or data privacy — self-hosting wins decisively.
Run your own numbers with the cost calculator.
What you can prepare today
- 1. Prototype against the hosted API. The fal endpoints expose the same modes (T2V, I2V, editing, reference-to-video) you'll run locally — build your product logic now, swap the backend later.
- 2. Sort your CUDA environment. Python 3.10+, CUDA 11.8+, recent PyTorch, and FP8 support if your card has it. Most self-hosting pain is environment pain.
- 3. Watch the release, not the hype. Our status tracker records exactly which artifacts are verifiably published, with dates.