Skywork, open-sourced 2026-01-29: three tasks, three weight sets — 14B reference-to-video, 14B video extension and 19B audio-driven talking avatar, all at 720P / 24fps / 5–30 sec. The README gives no absolute VRAM figure, only a line in the sand: under 24GB you must pass --low_vram (FP8 weight-only quantisation plus block offload) and drop to 540P / 480P. Single-GPU inference works; multi-GPU uses xDiT USP sequence parallelism (the official example runs 4 cards). It accepts 1–4 reference images, one of the few open models here that can lock a subject.
Weights are free, hardware is not: you buy the GPU up front (one-time estimate, may rise) and still pay for power and time. Six of the twelve are Apache 2.0 (HappyHorse-1.0, SANA-WM, Bernini, MAGI-2, LingBot-Video, LongLive 2.0); LTX-2 and LTX-2.5 use community licences and SANA-Video uses the NVIDIA Open Model License. Hardware runs from one RTX 5090 to 8× H100. See all 12 →
For GPUs with lower VRAM (e.g., under 24GB), use these options — --low_vram enables FP8 weight-only quantisation plus block offload, and you drop the resolution to 540P / 480P.SkyReels-V3 GitHub README
| # | Platform / model | Layer | VRAM · setup | One-time hardware |
|---|---|---|---|---|
| — | OSSSkyReels-V3 ↗Skywork · 2026-01 · custom licence · 14B / 19B at 720P | Open source | 24 GBUnder 24GB needs FP8 + offload | ≈ ¥24,000 · one-time1× RTX 4090 24GB · estimate, may rise |