Ant Lingbo, open-sourced 2026-07-09 — the first MoE video foundation model aimed at embodied AI: 30B-A3B (30B total, ~3B activated per forward pass) plus a 1.3B dense variant, Apache 2.0. Do not be fooled by "A3B": activating 3B saves compute, not VRAM — the other 27B of expert weights still has to sit in memory. Measured peak is 83.75 GiB per card at CP8 without sharding, falling to 37.61 GiB with DiT-FSDP (about 17% slower). The 1.3B dense build runs on a single 48GB card. The official benchmark denominator is 480×832 / 5 sec / 24fps. This one predicts worlds for robots, not shots for film.
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 →
A3B saves compute, not VRAM — the other 27B of expert weights still has to sit in memory.Ant Lingbo · github.com/robbyant/lingbot-video + community measurements
| # | Platform / model | Layer | VRAM · setup | One-time hardware |
|---|---|---|---|---|
| — | OSSLingBot-Video ↗Ant Lingbo · 2026-07 · Apache 2.0 · 30B-A3B MoE | Open source | 80 GB84GB peak per card, 38GB with FSDP | ≈ ¥2,000,000 · one-time8× H100 80GB · estimate, may rise · cloud rental is more practical |