MAGI-1

MAGI-1 is Sand AI's open-source video generation model (released April 21, 2025, Apache 2.0), built on an unusual idea: instead of rendering a whole clip at once, it generates video autoregressively in 24-frame chunks, each conditioned on the ones before it. That makes it the rare model with true infinite video extension — chain chunks with different prompts per segment for scene transitions and long-form narratives — plus chunk-wise prompting for fine-grained control as a scene unfolds. It scored above Kling 1.6 and Wan 2.1 in community comparisons and leads on the Physics-IQ benchmark for realistic motion. A 4.5B distilled variant runs on a single RTX 4090; ComfyUI nodes exist.

The honest picture: the flagship 24B model needs 8x H100s — the 4.5B version is the realistic local option. Non-photorealistic styles lag behind its realistic output, and single-shot visual polish trails closed models like Sora/Kling on short clips. But for open, extendable, physics-plausible video — the world-model direction — it is genuinely distinctive.

Use the hosted platform at sand.ai, or self-host via Docker with weights from Hugging Face; ComfyUI custom nodes available since May 2025.