MAGI-1

Sand AI's MAGI-1 builds a clip piece by piece: video is cut into 24-frame chunks, and work on the next chunk starts while the previous one is still partly noisy, with up to four in progress together. Since every chunk depends only on earlier ones, output can stream as it is made, and the model can extend an existing video or take a new prompt per chunk for scene changes and longer sequences. One pipeline script handles text to video, image to video, and video continuation, running a diffusion transformer with a transformer VAE and a T5 text encoder. Published checkpoints include 24B and 4.5B base models, distilled and FP8 quantized variants, and MAGI-1.1 24B weights added in June 2026. The large models are demanding, targeting eight H100-class GPUs, while the 4.5B model runs on a single 24 GB card and its quantized distilled version can squeeze into 12 GB with a reduced window setting. A Docker image, a source install path, and ComfyUI nodes are available. Code and weights are Apache 2.0, which permits commercial use, making it a practical pick for researchers and studios exploring long or streamed video generation.