Video catalogue · Verified local path

MAGI-1 local guide

Apache-licensed autoregressive video generator for local text-to-video, image-to-video and video-continuation workflows.

Choose an app

Start with Recommended. No terminal commands are shown.

Compare video models
Open model filesRecommended verified starting pointOpen in UnslothLaunches Unsloth Desktop on this modelGet ComfyUI DesktopOfficial desktop app for this compatible modelOpen advanced setupPython runtime for experienced usersCheck my machineSee whether this model fits your saved hardware

Desktop app links require the app to be installed. If nothing opens, LocalClaw will show app-download and model-file fallbacks.

What it does

Text To VideoImage To VideoVideo To VideoVideo ContinuationStreaming Video

Sand AI publishes Apache-2.0 code, Docker/source install instructions, ComfyUI support and Hugging Face weights for MAGI-1, including 24B, 24B distilled/quantized, 4.5B and MAGI-1.1 24B variants. The official model zoo recommends one RTX 4090 for the 4.5B checkpoint and states that 4.5B models need at least 24 GB GPU memory, while the 24B line targets four to eight H100/H800 or RTX 4090 GPUs depending on quantization. LocalClaw records 64 GB RAM and 24 GB NVIDIA VRAM as the conservative single-GPU floor for the 4.5B local path.

Hardware figures are practical entry floors, not performance guarantees. Resolution, duration, precision, offloading and runtime versions can materially change memory use.

Strengths

  • Official Sand AI code, Docker path and downloadable weights
  • T2V, I2V and V2V modes through the same MagiPipeline scripts
  • Autoregressive chunk generation for long-horizon continuation and prompt changes

Limits to know

  • Highest-quality 24B models require multi-GPU H100/H800 or many RTX 4090 cards
  • Linux/CUDA research stack with large model downloads
  • The 4.5B checkpoint is the practical local entry point, not the flagship 24B configuration