Zuckerberg wants more open-source AI: here's how closed models differ from open ones

Meta chief executive Mark Zuckerberg has renewed his push for fewer regulatory barriers around open-source artificial intelligence, arguing that the United States needs a more open AI ecosystem to compete effectively with fast-moving Chinese rivals.

The comments came as Meta unveiled Muse Glimmer, a new open-weight AI model designed to run directly on consumer devices rather than expensive cloud infrastructure.

Zuckerberg also indicated that Meta intends to release more models in the coming months, reinforcing its strategy of making AI more widely accessible.

The launch comes at a time when Meta is trying to regain momentum in the AI race after assembling a dedicated superintelligence team last year, while businesses increasingly search for cheaper AI alternatives amid soaring computing costs.

At the same time, the discussion has highlighted an important distinction that is often misunderstood — the difference between closed AI models, open-weight models, and fully open-source AI systems.

Silicon Valley has long been divided over how advanced AI should be developed.

Some companies argue that powerful AI models should remain closed because of safety concerns, while others believe openness is essential for innovation and competition.

Understanding that debate requires first examining what closed models are, then open-weight models, before finally looking at what constitutes true open-source AI.

What are closed models?

Artificial intelligence models learn by identifying patterns across enormous amounts of digital information.

As they train on books, news articles, websites and other online documents, they convert those patterns into billions of numerical values known as "weights."

These weights ultimately determine how an AI model reasons, responds to prompts and generates content.

The AI systems most consumers interact with today — OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini — are examples of closed models.

While users can access these systems through websites or application programming interfaces (APIs), the companies behind them do not release the underlying weights that power the models.

That means developers cannot download them, install them on their own computers, inspect how they work internally, or customize them for specific applications.

Anthropic and OpenAI have argued that frontier AI systems should remain tightly controlled because increasingly capable models present significant security risks and should be developed inside carefully managed environments.

Meta, Nvidia, Microsoft and Google have generally taken a more open stance, arguing that broader access encourages innovation and allows developers to build new businesses around AI technology.

Even so, researchers caution that closed models create their own risks.

According to Stanford University's Institute for Human-Centered AI, closed models are not necessarily less safe than open alternatives, but they concentrate enormous power within a handful of companies.

With frontier AI capabilities delivered through APIs, those companies decide who gains access, how much it costs, and which values become embedded in technology that increasingly shapes how people work, communicate, and create.

What are open-weight models?

Open-weight models occupy the middle ground between fully closed systems and completely open-source AI.

When developers describe an AI model as open-weight, they mean the model's trained weights have been made public.

That allows outside developers to download the model, run it locally, and fine-tune it for specific purposes, such as healthcare, cybersecurity, or software development.

If the weights remain private, users must interact with the AI exactly as its creators intended.

However, open-weight models do not reveal everything.

Companies typically release only the trained parameters while withholding the original training code, datasets, model architecture details, and much of the methodology used to build the system.

As a result, users gain flexibility without complete transparency.

Open-weight models also tend to be considerably cheaper than frontier AI systems offered by companies such as OpenAI and Anthropic because they can often run locally without expensive cloud computing resources.

Although several American companies publish open-weight models, many of the most widely adopted systems currently come from China, including Alibaba's Qwen family, DeepSeek's models, Moonshot AI's Kimi lineup and Zhipu AI's GLM series.

Their popularity has grown rapidly as businesses seek lower AI operating costs while retaining greater control over deployment.

Now, what are open-source models?

Zuckerberg's repeated calls for open-source AI have also revived debate over whether today's so-called "open" AI models are actually open source.

There is an important distinction between releasing an AI model with open weights and releasing it as fully open source.

Open weights involve publishing only the pretrained parameters that determine how the neural network behaves.

This enables developers to perform inference and fine-tune the model for downstream applications.

However, crucial elements remain unavailable, including the training code, original datasets, model architecture, and the process used to train the model.

Releasing AI in this manner broadens access but limits transparency, reproducibility and independent verification.

Developers using open-weight models must ultimately trust the judgments and representations learned by the original creators without being able to recreate or fully audit them.

A truly open-source AI model goes much further.

It includes everything required to reproduce the model from scratch — source code, architecture, training methodology, hyperparameters, datasets and supporting documentation.

That level of openness allows researchers not only to run the model but also to inspect it, improve it and build entirely new systems from the same foundation.

"There’s a wide gap between open-weight AI and open source AI," said Stanford Institute for Human-Centered AI Denning Director James Landay.

"Open weights answer 'Can I run this?' Open source answers 'Can I trust this, improve it, and build the next thing on top of it?' Right now almost everyone – American labs and Chinese labs alike – is answering the first question but nowhere close to the second."

Meta's push for open-source models

Alongside the model release, Zuckerberg published a 14-page essay outlining his vision for the future of artificial intelligence.

In the document, titled The Future is for Everyone, he argued that AI should be viewed as a tool for broad societal benefit rather than as a technology that must be tightly restricted.

He also stressed that the United States must maintain its leadership over China in the global AI race.

Without naming competitors such as Anthropic and OpenAI, Zuckerberg criticized approaches that advocate stricter controls over advanced AI systems.

He argued that concentrating the technology in the hands of a handful of companies could itself create risks.

“Rather than centralizing superintelligence, we should distribute it,” Zuckerberg, 42, wrote in the essay. “This has the potential to begin a new era of personal empowerment.”

The company was among the earliest major AI developers to publicly release a large language model, LLaMA, in February 2023. Meta published both the inference code needed to run the model and the weights learned during training.

However, organizations such as the Open Source Initiative have argued that LLaMA's licensing restrictions — particularly those governing commercial reuse — prevent it from meeting the accepted definition of genuine open-source software.

Instead, many experts classify it as an open-weight model.

Still, many developers believe that genuine open-source AI should include not only code and weights but also the datasets used during training, allowing complete transparency into how these increasingly influential systems are built.