A China -based startup, Sand AIMicrosoft Research Asia’s founding director Kai-Fu Lee has published a public licensed video-maker AI model who has received praise from entrepreneurs from entrepreneurs. According to the TechCrunch test, Sand AI is censoring images that can increase the concerns of Chinese regulators from the hosted version of the model.
In the beginning of this week, Sand AI has announced Magi -1A model that creates videos by “Autorgressly” by predicting the frame sequences. The company claims that the model can create high quality, controllable footage that captures physics more accurately than rival exposed models.
đź‘€ You don’t believe it is AI-—
🧠You don’t believe it is open source
🎬 You don’t believe it is freeMagi-1 video model has simply insulted commercial video equipment
Below details and examples: 👇 pic.twitter.com/zlxrecweqh
– Farhan (@MHDIFIRN) April 21, 2025
Maggie -1 is very unreasonable for most consumer hardware. It requires 24 billion parameters in size and four to eight nvidia H 100 GPU to run it. (Parameters are using internal variable models to predict)) For many users-this reporter includes the platform of the Sand AI is the only place that they can test Drive Maggie-1.
Video generation requires a “prompt” image of the platform to close. Not all prompts are allowed, TechCrunch has been discovered quickly. Sand AI blocks uploads of insignius that supports Xi Jinping, Tiannman Square and Tank Man, Taiwanese flag and Hong Kong Liberation. Filtering seems to be happening at the figure level; Blocking skirts named after image files did not.
Sand AI is not the only Chinese startup that prevents politically sensitive images in its video generation equipment. Hailoo AI, Shanghai-based Minimax’s generator media platform, also blocked photos of Xi Jinping. However, filtering of Sand AI seems specially aggressive; The Hoilo also allowed the images of Tiananmen Square.
As ward explained in a piece from January, Chinese models need to follow strict information control. A 2023 Law models “damage the country’s unity” – it forbids the government’s historical tihassic and political narrative to create content. To adhere to, Chinese startups often censor their models, either through prompt-level filters or fine tunes.
Interestingly, although Chinese models blocked political lectures, they often have less filters than their American part for pornographic materials. 404 has recently been reported Several videos published by Chinese companies lacks basic maintenance of generators that prevent people from creating non -sensitive nudity.