Zuckerberg admits "Llama 4 wasn't it" as Meta open-sources Muse

Admitting the Llama 4 misstep, in his own words

Mark Zuckerberg reportedly made self-critical remarks about Llama 4 in an interview with YouTuber and newsletter writer Rowan Cheung. The published post shows a sentence cut off starting with "Llama 4 was not," so the exact full context hasn't been confirmed, but Rowan Cheung introduced it by saying "Meta admitted where it went wrong." Llama 4 is one of Meta's large language model (LLM) lines, a multimodal model that handles both text and images. Since its release last year, it has been the subject of controversy over benchmark scores and its standing within the open-source community, and this remark appears to be an extension of that ongoing discussion.

Muse Glimmer open-sourced, Muse Spark 1.2 coming soon

In the same interview, Zuckerberg said Meta would open-source the weights of its new model, "Muse Glimmer." Releasing weights means making the parameter values a model acquired through training publicly available as-is, allowing anyone to download and run it on their own servers or devices. He also teased that a follow-up, "Muse Spark 1.2," would arrive soon. However, the excerpted material does not include details on the parameter scale, specific performance, or licensing terms of either model, so further confirmation is needed.

Manifesto: "The Future is for Everyone"

Images released alongside the interview showed part of a document apparently written by Zuckerberg titled "The Future is for Everyone," subtitled "A path to a positive AI future." The document proposes a philosophy centered on individual empowerment as the source of prosperity, invention as the core purpose of superintelligence, and a balance of power as the foundation of safety. Notably, it includes a line to the effect that "I can't understand people who believe AI will eliminate most jobs and humanity's reason for existing, yet still rush to bring about that future." It also states opposition to the argument that extreme concentration of power is the only safe path forward.

This document reads as part of the same message as the 6,500-word essay Zuckerberg published on August 10. At the time, TechCrunch pointed out that the essay actually fueled public distrust of AI, and it was widely discussed alongside the fact that, that same weekend, a court fined Meta $567 million over matters related to harm to children. This interview comes at a point when that controversy has not yet fully subsided, which leaves room to interpret it as Zuckerberg's own attempt to explain a course correction by directly addressing the misstep.

Doubling down on open source

In recent months, Meta has continued to position itself against the closed camp represented by OpenAI and Anthropic on open-source strategy. Google, NVIDIA, and others are also joining the trend of using open-weight models as leverage to expand their own hardware and cloud ecosystems. This release of Muse Glimmer can likewise be read as a move to reclaim the open-source leadership that had been shaken by Llama 4.

So what changes now

It hasn't been confirmed whether this announcement alone means developers or general users can immediately download and use Muse Glimmer. But the fact that Meta itself pointed out Llama 4's flaws and is once again playing the open-source card is itself a signal. At a time when open-source models are rapidly catching up to closed ones in performance, Meta's next moves are set to be a bellwether for where leadership in the open-weight ecosystem heads after the Llama series.