SkyReels targets long-form generation specifically, where most other models here cap out at a few seconds. Three generations shipped since February 2025: V1 (human-centric), V2 (long-form), V3 (multimodal reference input).
SkyReels
by Skywork AI
Open-source video model family for long-form generation, self-hosted rather than API-only.
- Fully open-source: weights and code downloadable for local deployment, no per-call API fees required
- SkyReels-V2 (April 2025) uses a 'Diffusion Forcing' architecture for long/unlimited-length video; ships in 1.3B, 5B, 14B sizes
- SkyReels-V3 (Jan 2026) adds multimodal input — 1-4 reference images for characters/objects/backgrounds
- Also available via a web playground and third-party API resellers, in addition to self-hosting