Skywork released and open-sourced this on 2026-10-15 — code, weights and the white-box build. It is not a general text-to-video model but "the engine owns the world, the model paints it": a game engine keeps world state and rules, and the video model renders the engine’s structural frames into detailed visuals in real time. A single RTX 5090 (32GB) runs it at 480p / 60fps with a measured 18.5 GiB VRAM peak; key-to-first-frame latency is 112ms (official claim <150ms) and steady-state denoising reaches 83fps. It is Wan2.2 5B turned into a V2V model with 148-channel concatenated input. For finished 1080P footage you still want something else — the selling point here is real time, not image quality.
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 →
A single NVIDIA RTX 5090 runs the game at 60 FPS at 480p; key-to-first-frame latency is under 150ms; steady-state single-step denoising reaches 83 FPS with an 18.5 GiB VRAM peak.Skywork release note · 2026-10-15
| # | Platform / model | Layer | VRAM · setup | One-time hardware |
|---|---|---|---|---|
| — | OSSMatrix-Game 4.0 ↗Skywork · 2026-10 · code + weights + white-box · Wan2.2 5B base | Open source | 18.5 GBMeasured peak · one 5090, real time | ≈ ¥35,000 · one-time1× RTX 5090 32GB · estimate, may rise |