RepoPicks Radar

Updated every day around 23:30 UTC

Guides

MAGI-1

SandAI-org/MAGI-1Reviewed by AI

Checked by RepoPicks Radar on

What is MAGI-1?

MAGI-1 is Sand AI's video generation model with the code to run it and the model files published on Hugging Face. From the command line you pick a mode (text to video, image to video, or continuing an existing clip), write a prompt and get a video file; custom nodes also let it run inside ComfyUI. It is for people with serious hardware: the small 4.5B model asks for an NVIDIA card with 24 GB (12 GB in the quantized version) and the 24B model for eight H100 cards. The authors also offer MAGI-1 to use online at sand.ai.

MAGI-1 is free and open source (Apache-2.0) and can replace Runway, Pika, and LTX Studio.

How easy is it to install MAGI-1?

D · terminal, several tools or a specific graphics card · Linux

Linux is the only documented system and it is a D: the recommended path is `docker pull sandai/magi:latest`, otherwise conda with Python 3.10.12, PyTorch 2.4 with CUDA 12.4, ffmpeg and a build of MagiAttention from source, then editing a JSON config and running example/4.5B/run.sh. There are no instructions for Windows or macOS.

Coming soon: install in 1 click with the RepoPicks app.

MAGI-1 on GitHub

Stars
3,805on Nov 10, 2026
Downloads
—all releases
Growth
+7stars in 6 days

Stars over time

3,8033,8043,8053,806 Nov 6Nov 7Nov 8Nov 9Nov 10 3,8033,8043,8053,806 Nov 6Nov 8Nov 10

Facts

Runs on
Linux
You need to install
Python, Docker, NVIDIA graphics card, FFmpeg
To build from source
Python, FFmpeg
Disk space
150 MB–450 MB · estimate
Accounts or keys
None
Cost
Free · Code and model files are free under Apache-2.0; the real cost is the graphics hardware (or rented machines) needed to run them.
License
Apache-2.0 (Permissive)
Last change on GitHub
Jun 17, 2026

MAGI-1 risk signals (6)

  • AttentionDependencies with known vulnerabilities

    Pinned dependencies with known vulnerabilities (OSV): accelerate: GHSA-4j2p-28q2-5m79, PYSEC-2026-3804; diffusers: GHSA-7wx4-6vff-v64p, GHSA-98h9-4798-4q5v, PYSEC-2026-2446; flash-attn: GHSA-7g5w-pq96-8c5w, PYSEC-2026-2477; protobuf: GHSA-7gcm-g887-7qv7, GHSA-8qvm-5x2c-j2w7, PYSEC-2026-1805; sentencepiece: GHSA-38vq-g6vr-w8wf, PYSEC-2026-1909; transformers: GHSA-29pf-2h5f-8g72, GHSA-37mw-44qp-f5jm, GHSA-4w7r-h757-3r74

    automatic check

  • Attentionneeds expensive graphics hardware

    Model Zoo recommends "H100/H800 x 8" for the 24B models and "RTX 4090 x 1" for the 4.5B ones; the note says "any machine with at least 24GB of GPU memory is sufficient" for 4.5B, or 12 GB with the quantized model and window_size set to 1.

    AI review

  • Attentiondocumented docker run is privileged

    The recommended command is `docker run -it --gpus all --privileged --shm-size=32g --net=host --ipc=host --ulimit memlock=-1 ... sandai/magi:latest`: the container runs with privileged access, the host network and the host's shared memory.

    AI review

  • Infoinstalls from an extra package index

    requirements.txt line: `flashinfer-python==0.2.0.post2 --extra-index-url https://flashinfer.ai/whl/cu124/torch2.4/`, so pip also pulls a package from flashinfer.ai instead of only from PyPI; the source path for MagiAttention is `git clone [email protected]:SandAI-org/MagiAttention.git`, which needs an SSH key on GitHub.

    AI review

  • Infomodel files come from hugging face

    Model Zoo links every checkpoint (T5, VAE, 4.5B and 24B, distill and fp8_quant) to huggingface.co/sand-ai/MAGI-1; the repository itself ships only code, so the weights and their license terms are downloaded separately.

    AI review

  • Infono releases and no generation mark

    The project facts say "GitHub releases: 0" (changes are announced only in the README's news list, latest 2026-06-17), and nothing in the README mentions a watermark or any mark on the generated videos.

    AI review

Evidence is shown in its original language.

What Radar could not check about MAGI-1

  • what the online service at sand.ai costs, since the README only says "directly create magic with MAGI-1 [here](http://sand.ai)" and mentions no price or plan
  • how much disk the model files really take, since they are downloaded from Hugging Face and are not part of the repository
  • how long one video takes on a single RTX 4090 and what resolution and length are usable there
  • whether the published videos and evaluation scores can be reproduced at home
  • what license the model files on Hugging Face carry, which can differ from the Apache-2.0 code
  • whether the ComfyUI nodes still work with current ComfyUI versions

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub