Zuckerberg's Open-Source AI Pledge Unravels as Meta Goes Closed

The manifesto and the model

In July 2024, Mark Zuckerberg published an essay titled "Open Source AI Is the Path Forward" on the same day Meta released Llama 3.1. That release included a 405-billion-parameter model, which Meta called the first frontier-level open source AI model, along with improved 70B and 8B versions.11 He compared AI to the history of Unix and Linux. In his telling, the closed models of the day would give way to open ones, and he predicted that Llama models would become the most advanced in the industry starting the following year.1115 He called the release an inflection point, after which most developers would rely mainly on open source.26

Two years later, that prediction looks very different. In April 2026, Meta shipped Muse Spark, its first major proprietary model, built by the newly formed Meta Superintelligence Labs. There are no downloadable weights, and access runs through Meta AI and a private API preview.110 Going back to the manifesto now shows how much of Meta's open-source case depended on circumstances, and how exposed that left the developers who believed it.

What Meta actually shipped in 2024

The technical case was strong. The 405B model was trained on more than 15 trillion tokens using over 16,000 Nvidia H100 GPUs. The whole Llama 3.1 family got a 128,000-token context window and support for eight languages.1621 Meta published benchmarks showing the model competing with GPT-4, GPT-4o and Claude 3.5 Sonnet.12 Ars Technica described it as possibly the first time anyone could download a GPT-4-class model for free and run it on their own hardware. The outlet also warned that standard benchmarks tell users little about how a model feels to work with.16

"Run it yourself" came with a large asterisk. Meta said the model could run on a single server node, which is far beyond desktop hardware.16 Notebookcheck reported that local inference needed roughly 810 GB of GPU memory spread across multiple high-end accelerators.27 Fortune pointed out that the model was too large for a single GPU, or even a dozen.23

The developer tools play

This is the part of the launch that mattered most for working engineers. Zuckerberg said Meta had built Llama models mainly for itself and then released them, without investing much in the wider ecosystem. Llama 3.1, he wrote, would be different, with internal teams and partnerships set up to help developers adopt it.11

That promise was real at launch. The models were available from more than 25 partners on day one, including AWS, Nvidia, Databricks, Groq, Dell, Azure and Google Cloud. Meta also announced support for vLLM, TensorRT and PyTorch so developers could start building right away.22 Meta released the Llama Stack specification for building agentic applications, plus the Llama Guard 3 and Prompt Guard safety classifiers.21 Weights could be downloaded from Meta, from Hugging Face, or through Ollama.28

The most consequential change was in the license. For the first time, Meta officially allowed developers to use Llama outputs to improve other models.16 Zuckerberg presented the 405B model as a teacher for fine-tuning and distilling smaller models, and claimed that running it on your own infrastructure would cost about half as much as using closed models like GPT-4o.11 Ahmad Al-Dahle, Meta's VP of generative AI, said he expected developers to use it in exactly that role.26

Not everyone accepted the cost claims. Fortune noted reports that some developers had found self-hosting Llama 3's 70B model more expensive than paying OpenAI per token for GPT-4. It also questioned how much control customers would have over model weights running on partners' shared infrastructure.23

"Open source" was disputed from day one

The launch coverage split sharply on terminology. Meta's own materials and much of the enthusiast press called Llama 3.1 open source with no qualification.1314 Critics did not let that stand.

The Open Source Initiative's executive director, Stefano Maffulli, told ZDNet that the license still restricted which companies could use the model, which made it "clearly and inarguably not open source." He also called Llama a liability for developers, saying the license ultimately left Meta in charge of their innovations.22 A Linux developer noted that Fedora had been unable to ship Llama 2 because its license was not OSI-approved.22 Nathan Lambert of the Allen Institute for AI accused Meta of "open-washing," saying the release documents described training data only as "publicly available," without definition.25 Ars Technica pointed out that downloading the model required agreeing to a license and acceptable use policy, which in principle let Meta revoke access. The outlet chose to use the term "open weights" instead.16

The license terms supported the critics. It required Meta's approval for companies with more than 700 million monthly active users. It also required distilled models that are distributed to carry "Llama" at the start of their names.21 Some commentators argued that Meta's interest in making AI models a commodity matched its advertising business more than any principle.21 Zuckerberg was fairly open about the self-interest. He framed open releases as a way to avoid paying tolls to platform owners, comparing it to the fees Apple charges developers through its App Store.16

One point from the time now reads like a warning. Zuckerberg argued that because AI competition would stay intense, open-sourcing any single model wouldn't hand rivals a big advantage. Llama's path to becoming the standard, he said, depended on staying competitive generation after generation.18 Put another way, the openness pledge depended on Llama staying near the top.

Why the pledge broke

It didn't stay near the top. Reports in December 2025 said Meta was developing a proprietary model codenamed Avocado after the internal failure of Llama 4 "Behemoth."5 Coverage also described internal frustration that Llama 4 had not excited developers and that rival models had reused parts of Llama's architecture.9 Zuckerberg had already started hedging, writing that Meta needed to be "careful about what we choose to open source."59

The stated reasons differ in emphasis. One analysis points to three factors working together. First, a benchmark controversy around Llama 4, in which Meta tested an unreleased experimental chat version on LMArena, damaged its credibility. Second, Meta's investment in Scale AI created a proprietary data pipeline. Third, competitors could use open weights to train copies of Meta's models.1 Other reporting puts more weight on Llama 4's lukewarm reception and on worries that labs such as DeepSeek were building on Meta's open research.7 Meta has not given a full explanation, and it did not respond to The New Stack's questions on the subject.10

My reading is that the competitive argument mattered more than any of these individual factors. The manifesto assumed open weights would lock in Meta's position as the standard. Once Llama stopped leading, openness mostly helped rivals, and the strategic case for it collapsed.

Where developers stand now

Meta's public line is that current Llama models will remain available as open source. That commitment covers existing models only and says nothing about future development.10 Zuckerberg and Alexandr Wang have both mentioned open-sourcing future models, but neither has given a timeline or a specific commitment.1 One outlet reported in April 2026 that Meta had confirmed open-source versions of its next generation, also without naming models or dates.2 Some descriptions say Muse Spark ships alongside Llama rather than replacing it.8 Others say Meta has effectively abandoned new Llama development.10 The difference is mostly about framing. Both accounts agree that Meta's frontier work is no longer open by default.

The practical consequences are clear. The New Stack notes that there is no migration path from Llama to Muse Spark, because one offers self-hosted weights and the other is cloud-only.10 Andrew Ng called Meta's move away from being the leading US champion of open weights "a significant loss for the developer community."10 Developers can stay on Llama models that will fall further behind, move to open alternatives from Mistral, DeepSeek or Qwen, or switch to proprietary APIs.10

The open ecosystem itself is not shrinking. By one count, the open-source community published more than 85,000 Llama derivatives on Hugging Face in 2025 alone.4 Competitive open-weight models now come from Google, Alibaba, Zhipu, DeepSeek and Mistral as well as Meta.1

That may be the most lasting result of the 2024 essay. It helped show that open-weight models could compete at the frontier, and that argument now carries on without Meta. The critics who disputed the "open source" label in 2024 were right about one thing: a license controlled by a single company left that company in charge of what developers could build.22