With new open models, Meta pitches another reboot of its struggling AI strategy - Ars Technica

Overview

With new open models, Meta pitches another reboot of its struggling AI strategy

Meta has been trailing competitors. Zuckerberg thinks he’s found a way forward.

Details

Meta has announced its intention to focus on open-weight large language models. Additionally, the company announced the release of an open model called Muse Glimmer and a promise to open the weights for Muse Spark 1.2, its more powerful model, in the next few weeks.

Alongside these announcements, Meta CEO Mark Zuckerberg published a more than 6,000-word essay outlining the company’s philosophy about AI systems and governance moving forward. The essay aims to differentiate Meta from companies like Open AI and Anthropic, which develop proprietary models and which have lobbied the US government for help competing against large-scale distillation—which involves using an existing model to train a new one—or open-weight models by Chinese labs.

Muse Glimmer is a 30 billion parameter model with a 128,000-token context window by default. It is distilled from Muse Spark, the larger and more capable model that Meta launched earlier this year. Glimmer is meant to run on users’ local machines, rather than via a cloud service or an API. Glimmer’s weights are open source under the Apache 2.0 license.

Muse Spark was introduced in April as a closed, proprietary, frontier-class model—Meta’s first major model release after a significant shake-up of the company’s AI teams last year, and a departure from its focus on models that are, by some definition, open. When Meta released Muse Spark 1.1 in July, it introduced its first paid service—again, a departure from its previous strategy. Muse Spark 1.2 was released on August 5 and was accompanied by Muse Code, a terminal coding agent.

Developers have generally found that Muse Code doesn’t quite match the frontier models from Anthropic or Open AI in capability, but it competes well on cost—meaning it has similar positioning to many open-weight models from Chinese labs.

As a smaller model designed to run on consumer GPUs, Muse Glimmer won’t compete on that level at all—but it reflects a growing movement to bring some inference to local devices to reduce reliance and spending on the models produced by the big labs like Anthropic and Open AI.

Alongside the model releases, Mark Zuckerberg is credited as the author of a lengthy open letter that details Meta’s corporate strategy with AI, and a broader argument for how AI should be developed, distributed, and regulated.

It follows several statements by other Big Tech and AI company leaders debating the merits of open-weight models, proprietary labs, and the practice of distillation—something that some Chinese labs have reportedly done to build models that compete with the latest efforts from Anthropic and others.

On July 24, several companies including Nvidia, Hugging Face, Meta, Mistral, Mozilla, Open AI and others co-signed an open letter titled “Open Weights and American AI Leadership” that argued for the value of open-weight models as opposed—or at least in addition—to proprietary ones, and defended distillation as a legitimate practice, even as it carved out a distinction for “unlawful efforts to extract value from closed models.” It advocated for “targeted legal and commercial frameworks rather than sweeping restrictions on techniques that play an important role in AI innovation.”

On the topic of distillation, Meta and Zuckerberg wrote:

The ability for models to learn from other models is an important principle of how the open source ecosystem works. All AI models are derived from human knowledge. Some have tried to frame distillation as harmful, but I think it is important to protect the principle that you can learn from anything you can observe.

However, the majority of the essay focused on the notion of decentralized AI systems that are distributed widely. It also directly aims at arguments that tightly controlled, proprietary AI models and systems are needed because of concerns about existential threats or catastrophic misalignment:

It is surprising that the discourse from many developing AI is so filled with doom. I do not understand why anyone who believes that AI will eliminate most jobs and much of humanity’s relevance would rush to build that future. The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic.

Zuckerberg wrote that the view of alignment taken by companies like Anthropic—which claims that it is trying to train singular and broad foundation models with operational parameters and guardrails that will ensure they benefit humanity broadly—is “fundamentally flawed.”

“People’s diverse values represent different tradeoffs they would make on important issues. There is no technological solution that can align with everyone’s opposing interests and values at once,” the essay says. “Any singular superintelligence would have to prioritize some values over others and in the process would be incapable of being benevolent to everyone.”

Instead, Meta argues here that models should be personalized to the needs and values of individuals or groups of individuals. It also claims that decentralization will make everyone safer, because it will give the benefits and advantages of “superintelligence” to everyone equally, instead of privileging “a small number of individuals, businesses, governments, or AI itself.”

Meta has been lagging behind other big tech companies and major frontier labs for foundation models. Its models haven’t seen the kind of adoption that those developed by Open AI or Anthropic have.

Open AI and Anthropic have aggressively targeted enterprise customers, releasing powerful models and harnesses for knowledge work tasks like software development, and they have made significant inroads and generated substantial revenue from this strategy. Meta has not seen the same level of success.

Meta also saw a total overhaul of its AI division last year, when former Meta AI chief scientist Yann Le Cun was replaced by former Scale AI CEO Alexandr Wang. The reset led to a change in focus.

In recent months, the debate around open-weight models and distillation has increased in volume as recent Chinese models like Alibaba’s Qwen 3.8-Max and Moonshot’s Kimi K3 have been shown to rival Anthropic and Open AI at the frontier. Those models may perform slightly worse in coding benchmarks, for example, but they are generally cheaper to use.

In a sense, Meta seems to be positioning itself as a US alternative to Alibaba, Moonshot, or Deep Seek—not quite as frontier-facing as Anthropic or Open AI, but more open, customizable, and affordable. It is also orienting itself—at least with these public statements—more toward personal use as opposed to large-scale enterprise deployments, at least for now.

That is a retreat from some of its earlier ambitions, in a way, as the company takes advantage of changing winds to try to plot a new course.

         Flesh-eating screwworms feast on humans in Mexico; human cases top 500

         Mount Toba eruption doesn't seem like it could nearly kill our species

         The first self-driving vehicle on Mars has proven to be a smashing success

         One of China's workhorse rockets just exploded in flight

         A researcher bought noreply.net. Companies started sending him secrets.

Ars Technica has been separating the signal from the noise for over 25 years. With our unique combination of technical savvy and wide-ranging interest in the technological arts and sciences, Ars is the trusted source in a sea of information. After all, you don’t need to know everything, only what’s important.

Key Takeaways

With new open models, Meta pitches another reboot of its struggling AI strategy
Meta has been trailing competitors
Meta has announced its intention to focus on open-weight large language models
Alongside these announcements, Meta CEO Mark Zuckerberg published a more than 6,000-word essay outlining the company’s philosophy about AI systems and governance moving forward
Muse Glimmer is a 30 billion parameter model with a 128,000-token context window by default