Meta has launched Muse Spark Open, a huge open-weight AI model with 816 billion parameters. It is built for coding, reasoning, and “agentic” work—where AI can take instructions and complete multi-step tasks. Meta calls it the most capable open-weight model from the United States, though Chinese rivals remain ahead on several coding tests.
This matters because the model’s weights are publicly available: developers can download, adapt, and run it rather than depending entirely on a company’s cloud service. It can understand text, images, audio, and video, and handle a context window of up to one million tokens. However, it is “open-weight,” not fully open source, because Meta’s licence imposes limited commercial conditions.
The Apple link is the most striking part. Meta and Apple have optimised the model to run locally on two Mac Studio desktops with M5 Ultra chips and 512GB of memory each, using Apple’s MLX software framework. In simple terms, two high-end Macs can be connected to act like a private AI computer —without sending every request to the cloud.
This does not mean ordinary users will soon run such a model at home; the hardware requirement is still extremely expensive. But it signals a shift: powerful AI may increasingly move from distant data centres to personal and office machines, improving privacy, speed, and control. Meta has re-entered the open-model race, giving the US a serious answer to China’s Qwen, GLM, and Kimi models.
AI IS LEAVING THE CLOUD—AND HEADING HOME.
Sanjay Sahay
Have a nice evening.