Meta, the company behind Facebook, Instagram, and WhatsApp, has released a new open-weight artificial intelligence model called Glimmer, alongside a sweeping 14-page essay by CEO Mark Zuckerberg titled "The Future is for Everyone." The essay argues that AI should be widely distributed rather than controlled by a small number of powerful companies or governments, and calls on the United States to ease restrictions on open-source AI development in order to compete more effectively with China.
Glimmer is a smaller, more efficient model designed to run directly on consumer devices such as a Mac or PC with a single graphics card, making it well-suited for so-called "agentic" tasks — software that can carry out instructions without constant human oversight. Unlike fully open-source software, open-weight models make the core rules governing an AI system's behaviour publicly accessible and customisable, while not necessarily exposing all underlying code. Meta also announced plans to release the weights of Muse Spark 1.2, its most advanced model, developed by an internal superintelligence team the company assembled last year. Zuckerberg said even larger models are forthcoming.
The release sharpens an already significant divide in the AI industry. Rivals OpenAI — maker of ChatGPT — and Anthropic, maker of Claude, focus on closed AI models that users cannot inspect or replicate, arguing that the rapid pace of AI development and the potential for misuse, including in cyberattacks or the development of biological weapons, makes tight control essential. Zuckerberg pushes back forcefully, writing that "the notion AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic." He also criticised competitors for building AI primarily for governments and large institutions, warning this would tilt power away from individuals. Notably, both OpenAI and Anthropic hold lucrative US government contracts.
The competitive pressure from China looms large over the debate. Chinese developers including DeepSeek, Alibaba, and Moonshot have in recent months released advanced, affordable open-weight models that rival the performance of leading US systems, while America's frontier labs have largely kept their most capable models closed. Zuckerberg argued that American regulatory constraints — on training data use and on "distillation," a technique for training smaller models using outputs from larger ones — put US firms at a disadvantage, and warned that even a one-month delay in releasing models could erode American leadership. Meta plans to implement an internal governance structure giving independent directors oversight of safety criteria for model releases.
The announcement carries clear commercial dimensions. Meta's shares, which had fallen roughly 10% this year, rose nearly 3% in premarket trading on the day of the release. The company, which has 3.6 billion users across its platforms, expects to spend up to $145 billion on AI infrastructure in the current year, nearly double its 2025 spending. To address growing public opposition to data centre construction — which has become a notable political issue in the US — Meta announced a $1 billion community support fund for areas hosting its facilities. The essay and model launch also come as Meta faces mounting legal pressure, including jury selection beginning this week in a California trial brought by four US states seeking $1.4 trillion in damages over alleged harms to children using its platforms.