ByteDance, open-sourced 2026-06 under Apache 2.0. 14B parameters on a Wan2.2 T2V-A14B base. Its angle is "understand first, render second": a multimodal LLM plans the semantics, then a DiT renders — which is why it can edit a specific region (V2V, reference-guided edits, content insertion, 5-image R2V). The hardware bar is high: the official setup targets Hopper-class cards (H100 / H800 / H200) for FlashAttention-3; consumer cards are not realistic.
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 →
The official deployment path targets Hopper-class cards (H100 / H800 / H200) and relies on FlashAttention-3.ByteDance · huggingface.co/bytedance/bernini-r
| # | Platform / model | Layer | VRAM · setup | One-time hardware |
|---|---|---|---|---|
| — | OSSBernini ↗ByteDance · 2026-06 · Apache 2.0 · 14B | Open source | 80 GBOfficial: H100/H800 (FA3) | ≈ ¥250,000 · one-time1× H100 80GB · estimate, may rise |