Most of the AI race has been about building ever-bigger models in ever-bigger data centres. Meta just went the other way. On Monday the company released Muse Glimmer, an open-weight model designed to run AI agents locally, on a normal Mac or PC with a single consumer graphics card. It is a smaller, freely downloadable version of Meta's more powerful — and still closed — Muse Spark model.
Glimmer has 30-billion parameters, a fraction of the size of the frontier systems it competes with, and its weights are released under a permissive licence so developers can download, fine-tune and modify it. It is built to actually do things — calling tools, writing and debugging code, working with files and screenshots, and running multi-step tasks — rather than just chat, and it is designed to work "anywhere, anytime, with or without an internet connection". It handles text and images and was trained across more than 100 languages.
The bigger argument
The launch came wrapped in a roughly 6,500-word essay from Meta chief executive Mark Zuckerberg, titled "The Future is for Everyone". His argument: powerful AI, or "superintelligence", should be spread widely to empower individuals, not concentrated in a handful of institutions. Meta, he wrote, is the lab focused on building "personal superintelligence for everyone", and it will offer free versions accessible to billions of people. He also pressed Washington to lower barriers so American open models can outcompete Chinese rivals like Alibaba, Moonshot and DeepSeek, which currently lead the open-weight field.
Not everyone is convinced
The reaction was not all warm. The Verge dubbed the essay a case for "unleashing Pandora's box". Critics also pushed back on the label: "open-weight" means you can download and run the model, but it is not the same as fully open-source, and even Meta's releases fall short of that stricter bar. Meta's shares, down about 10% this year, ticked up 1% in premarket trading on the news. For South African developers and small businesses, though, the real headline is simpler: a capable AI agent you can run on hardware you already own is a very different future from renting one from a data centre.
