Meta announced it will open source its latest and most powerful AI model, making its weights publicly downloadable, while also introducing a new family of lightweight models designed to run on laptops. The move, detailed by CEO Mark Zuckerberg in an Instagram video and a 6,500-word essay published Monday, positions Meta against closed-model leaders OpenAI and Anthropic, as well as Chinese open-weight competitors like Alibaba, DeepSeek, and Moonshot.
Zuckerberg said the company will release Muse Spark 1.2 with open weights, referring to the internal calculations that govern AI behavior, and unveiled the Muse Glimmer family for consumer devices. Shares rose 2.1% in premarket trading Monday, though they remain down about 10% this year amid investor scrutiny over Meta's spending, including capital expenditures projected at up to $145 billion for 2025. Zuckerberg aimed to reassure investors that Meta Superintelligence Labs, formed last year, is making progress.
In his essay, Zuckerberg urged U.S. policy changes to reduce restrictions on training data, arguing that foreign labs currently hold advantages because American firms face additional compliance hurdles. He rejected restricting access to foreign open models, stating the goal should be for American open-source models to lead globally. Counterpoint Research co-founder Neil Shah told CNBC that developers may pivot to Chinese open-weight models if Western tech giants build only walled gardens, and that Meta can fill the demand for non-Chinese open models.
Muse Glimmer's on-device design, processing AI locally rather than in cloud data centers, could cut costs and improve speed on consumer hardware, Shah noted, allowing Meta to outcompete Google and Microsoft on end-user devices. Zuckerberg also criticized the concentration of AI power among a few companies, appearing to swipe at OpenAI and Anthropic, whose CEOs have warned about AI's impact on jobs. He advocated for distributing superintelligence widely, envisioning personal agents that understand individual users and their goals. The essay also touched on distillation, the practice of using one model's output to train another, which some U.S. lawmakers view as intellectual property theft.
Source: CNBC