Meta is stepping up its push into open-weight artificial intelligence, with CEO Mark Zuckerberg announcing plans to release the company’s most powerful models publicly and introduce a new generation designed to run directly on consumer devices.
The strategy marks the boldest effort yet by Meta to differentiate itself from rivals such as OpenAI and Anthropic, whose leading models are largely developed and distributed as closed systems. It also reflects the effect of growing competition from Chinese AI developers, including Alibaba, DeepSeek and Moonshot, which have rapidly advanced open-weight models that developers can download, modify and deploy.
In an Instagram video on Monday, Zuckerberg said Meta would open the weights of its latest AI model, Muse Spark 1.2, allowing users and developers to download and use the system. Model weights contain the numerical parameters that determine how an AI system processes information and generates responses.
Register for the next Tekedia Mini-MBA.
Register for Tekedia AI in Business Masterclass.
Join Tekedia Capital Syndicate and co-invest in great global startups.
Meta will also introduce Muse Glimmer, a new family of open-source models designed to run on laptops.
The announcement comes as Zuckerberg faces mounting pressure to demonstrate that Meta’s enormous AI spending is producing meaningful technological gains. The company expects capital expenditure to reach as much as $145 billion this year as it builds computing infrastructure and expands its Meta Superintelligence Labs, which was formed last year.
Meta shares rose 2.1% in premarket trading on Monday but remain down about 10% this year, as investors weigh the scale of the company’s AI investment against the potential returns.
The open-weight strategy has the potential to give Meta a way to compete for developers and businesses without having to match the closed-model strategies of OpenAI and Anthropic on every dimension.
Chinese companies have already used open-weight AI as a competitive tool. Alibaba, DeepSeek and Moonshot have released models that have attracted international attention and, in some cases, competed with leading U.S. systems on specific tasks.
Zuckerberg used a 6,500-word essay published Monday to argue that the United States needs to make it easier for American companies to develop and distribute open AI models if it wants to maintain its lead over China.
“Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data,” Zuckerberg wrote. “US policy must reduce this additional friction if we want American open source models to lead over time.”
He also argued against restricting access to foreign open-weight models, saying the better strategy would be to make American models more competitive.
“I do not believe restricting access to foreign open source models is an effective solution,” Zuckerberg wrote. “Our goal should be for American open source models to be the best globally.”
The argument points to a difference between Meta and companies that favor tighter control over their models. Meta can use open weights to build an ecosystem of developers, enterprises and hardware manufacturers around its technology, potentially expanding the reach of its models without having to bear the entire cost of serving every AI query through its own data centers.
“If Western tech giants only build walled gardens, developers and enterprise builders will naturally pivot to Chinese open-weight models,” said Neil Shah, co-founder at Counterpoint Research.
“Most of its competitors in USA are proprietary and there is an insatiable demand for non-Chinese open models and weights and Meta can fill in this void well,” Shah said.
Meta’s second major bet is on running AI directly on devices.
Muse Glimmer is designed to operate on laptops, potentially reducing dependence on cloud-based AI services. Much of today’s generative AI processing occurs in data centers equipped with expensive GPUs and other accelerators. Moving some workloads to personal computers and smartphones could reduce cloud computing costs, improve response times, and allow certain AI functions to operate with less reliance on an internet connection.
“Bringing small, agentic models like Muse Glimmer directly onto PC and mobile hardware bypasses cloud compute costs to outcompete Google, Microsoft and others on the end-user’s device,” Shah said.
That strategy could become a dealbreaker as AI moves from simple chatbots toward agents capable of performing tasks on behalf of users. Smaller models that can operate locally could handle routine functions while larger systems remain in the cloud for more demanding workloads.
For Meta, the approach also creates an opportunity to connect AI more deeply with its enormous consumer ecosystem. Models that operate on personal devices could potentially support assistants, content creation, search, productivity and other functions without requiring every interaction to be processed remotely.
Zuckerberg’s policy argument extends beyond model weights.
He called for the United States to rethink rules surrounding data used to train AI systems and the practice known as distillation, in which outputs from a more capable model can be used to train another model. The issue has become contentious as AI companies and policymakers debate whether such techniques constitute legitimate model development or inappropriate use of another company’s intellectual property.
Zuckerberg also warned against concentrating control over advanced AI in a small number of companies, an argument that implicitly contrasts Meta’s open-weight approach with the closed strategies pursued by OpenAI and Anthropic.
“The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic,” he wrote.
His position places Meta at the center of a broader debate over whether increasingly capable AI should be controlled by a handful of companies or distributed more widely among developers, businesses and individuals.
The argument is also closely tied to the emerging debate over employment and the social consequences of increasingly capable AI. Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman have both warned about the potential effects of AI on jobs, although Altman has recently moderated some of his comments.
Zuckerberg instead framed widespread access to advanced AI as a mechanism for distributing economic and technological power.
“Rather than centralizing superintelligence, we should distribute it widely and give every person the ability to direct it,” he wrote.
“Everyone will have an exceptionally capable personal agent that understands you, your goals, and everything you care about,” Zuckerberg said.
The challenge for Meta is turning that vision into a commercially sustainable business. Open-weight models can accelerate adoption, but they can also make it harder for Meta to capture revenue directly from the underlying technology. At the same time, developing capable models requires enormous investments in computing infrastructure, data and talent.
That tension sits at the heart of Meta’s strategy. The company is spending up to $145 billion on capital expenditure while simultaneously arguing that the future of AI should be more open and distributed.