Sand AI, a China-based startup, has released an openly licensed video-generating AI model that has attracted praise from entrepreneurs like Kai-Fu Lee, founding director of Microsoft Research Asia. However, according to TechCrunch tests, Sand AI appears to be censoring images that could raise the outrage of Chinese regulators from the host version of the model.
Earlier this week, Sand AI announced the MAGI-1. MAGI-1 is a model that generates videos by predicting the sequence of frames “automatically over the net”. The company claims that the models can produce high quality, controllable footage that captures physics more accurately than rival open models.
👀I can’t believe this is generated by AI
I can’t believe it’s an open formula
🎬I can’t believe it’s freeA humiliating commercial video tool with just the MAGI-1 video model
Details and examples below: 👇pic.twitter.com/zlxrecweqh
– Fahan (@mhdfaran) April 21, 2025
MAGI-1 is so unrealistic that it cannot run on most consumer hardware. It has a size of 24 billion parameters and requires 4-8 NVIDIA H100 GPUs to run. (The parameters are internal variables that use the model model to create predictions.) Many users (which includes this reporter) are the only place where Sand AI’s platform can drive MAGI-1.
The platform requires a “prompt” image to start video generation. Not all prompts are acceptable, and TechCrunch was quickly discovered. Sand AI blocks image uploads of Xi Jinping, Tiananmen Square, Tank Man, Taiwan’s flag and insignia supporting Hong Kong’s liberation. Filtering appears to be happening at the image level. I renamed the image file and did not skirt blocking.
Sand AI is not just a Chinese startup that prevents uploading politically sensitive images to video generation tools. Shanghai-based Minimax’s generation media platform High-Ore AI is also blocking photos of Xi Jinping. However, Sand Ai’s filtering appears to be particularly aggressive. hauiluo allows images of Tiananmen Square.
As explained in wired in the January work, the Chinese model must follow strict information management. The 2023 law prohibits the model from generating content that “undermines the unity and social harmony of the country.” In other words, it contradicts the historical and political narratives of government. To comply, Chinese startups often censor models via prompt-level filters or fine-tuning.
Interestingly, Chinese models tend to block political speeches, but often have fewer filters than American counterparts in pornographic content. 404 recently reported that numerous video generators released by Chinese companies lack the basic guardrails that prevent people from generating unconsensual nudes.