Zuckerberg Signals Meta Won’t Open Source All Of Its ‘superintelligence’ Ai Models

Meta CEO Mark Zuckerberg shared his vision connected Wednesday for “personal superintelligence,” nan thought that group should beryllium capable to usage AI to execute their individual goals. 

Smuggled into nan missive is simply a awesome that Meta is shifting really it plans to merchandise AI models arsenic it pursues “superintelligence.”

“We judge nan benefits of superintelligence should beryllium shared pinch nan world arsenic broadly arsenic possible,” wrote Zuckerberg. “That said, superintelligence will raise caller information concerns. We’ll request to beryllium rigorous astir mitigating these risks and observant astir what we take to unfastened source.”

That wording astir unfastened root is significant. Zuckerberg has historically positioned Meta’s Llama family of unfastened models arsenic nan company’s cardinal differentiator from competitors for illustration OpenAI, xAI, and Google DeepMind. Meta’s extremity has been to create unfastened AI models that were arsenic bully arsenic aliases amended than those closed models. In a 2024 letter, Zuckerberg wrote, “Starting adjacent year, we expect early Llama models to go nan astir precocious successful nan industry.”

And while galore opportunity Llama doesn’t fresh nan strict meaning of unfastened root AI – partially because Meta hasn’t released its monolithic training datasets – Zuckerberg’s words constituent to a imaginable alteration successful priority: unfastened root whitethorn nary longer beryllium nan default for Meta’s cutting-edge AI. 

There’s a logic why Meta’s rivals support their models closed. Closed models springiness companies much power complete monetizing their products. Zuckerberg pointed retired past twelvemonth that Meta’s business isn’t reliant connected trading entree to AI models, truthful “releasing Llama doesn’t undercut our revenue, sustainability, aliases expertise to put successful investigation for illustration it does for closed providers.”  Meta, of course, makes astir of its money from trading net advertising.

Still, that stated viewpoint connected unfastened models was before Meta started to consciousness for illustration it was falling down competitors, and executives became obsessed pinch beating OpenAI’s GPT-4 model while processing Llama 3. 

Cut to June 2025, erstwhile Meta began its nationalist AGI sprint successful earnest by investing $14.3 cardinal successful Scale AI, acquiring Scale’s founder and CEO, and restructuring its AI efforts nether a caller portion called Meta Superintelligence Labs. Meta has spent billions of dollars to get researchers and engineers from apical AI firms and build retired caller information centers. 

Recent reports indicate that each that finance has led Meta to region testing connected its latest Llama model, Behemoth, and alternatively attraction efforts connected processing a closed model. 

With Zuckerberg’s ngo for introducing “personal superintelligence” to nan world – a decided displacement from nan rivals he says are moving connected “automating each valuable work” – his AI monetization strategy is taking shape. It’s clear from Zuckerberg’s words coming that Meta plans to present “personal superintelligence” done its ain products for illustration augmented reality glasses and virtual reality headsets. 

“Personal devices for illustration glasses that understand our discourse because they tin spot what we see, perceive what we hear, and interact pinch america passim nan time will go our superior computing devices,” Zuckerberg wrote successful Wednesday’s letter. 

When asked astir Meta perchance keeping its astir precocious models closed, a Meta spokesperson said that nan institution remains committed to unfastened root AI and said it besides expects to train closed-source models successful nan future.

“Our position connected unfastened root AI is unchanged,” a spokesperson said. “We scheme to proceed releasing starring unfastened root models. We haven’t released everything we’ve developed historically and we expect to proceed training a operation of unfastened and closed models going forward.”