Sand.ai (Cao Yue, an author of Swin Transformer) open-sourced this on 2026-08-05 under Apache 2.0. A 114B-parameter MoE that activates only ~6B per forward pass — the first open video model at the hundred-billion scale. Text, video and audio share one transformer, producing 10 seconds of 1080P with synchronised sound. It also carries the heaviest hardware bill here: about 228GB of BF16 weights, with cost estimates based on 8× H100, so individuals realistically rent it.
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 →
Roughly 228GB of BF16 weights, with cost estimates based on 8× H100.Sand.ai · github.com/SandAI-org/MAGI-2-preview
| # | Platform / model | Layer | VRAM · setup | One-time hardware |
|---|---|---|---|---|
| — | OSSMAGI-2 Preview ↗Sand.ai · 2026-08 · Apache 2.0 · 114B MoE | Open source | 8× 80 GB8× H100 · 228GB BF16 weights | ≈ ¥2,000,000 · one-time8× H100 80GB · estimate, may rise · cloud rental is more practical |