Meta's AI Chief Admits Open-Source Playbook Fails for Frontier Models

Meta's Chief AI Officer Alexandr Wang has admitted that the company's open-source AI playbook no longer fits its frontier models, with Muse Spark kept proprietary after early training flagged bio risk and other safety concerns. Wang says rival AI labs are seeing the same risks scale up. Meanwhile, Meta is testing subscriptions on Instagram, Facebook, WhatsApp and its AI chatbot to diversify beyond ads.

Alexandr Wang, the Scale AI co-founder Mark Zuckerberg hired last year as Meta's Chief AI Officer after a reported $15 billion deal, has acknowledged on Bloomberg Tech that Meta's longstanding open-source approach hit a wall with Muse Spark. The model, released in April, stayed proprietary because internal testing flagged risks the company couldn't safely contain in an open release, and Wang says rival labs are running into the same problem as their models scale. "It actually triggered some high risk areas in the course of early training, particularly around bio risk, but also a number of risks were elevated," Wang told Bloomberg. He added: "This is something I think the entire industry has seen as models improved dramatically over the past year."

The open-source bet quietly gets a rewrite

Meta built its AI reputation on Llama, the family of open-weights models that briefly made it the industry's standard-bearer for accessible AI. Wang's framing is more careful now. He says Meta will keep doing open source for models it judges "fit and safe" to release, while keeping frontier work locked down. Asked if Llama remains the brand for that effort, he sidestepped: "We have exciting debates about branding internally and nothing to share right now. " The pivot matters. As part of forming Meta Superintelligence Labs, Wang updated what he calls the company's advanced AI scaling framework, the internal document outlining how Meta evaluates and mitigates model risks. Muse Spark's in-product deployment, he argued, lets Meta apply guardrails that simply don't exist once weights are public.

Why Muse Spark still trails Claude and Gemini on coding

For all that recalibration, Muse Spark hasn't landed as a frontier challenger. The Financial Times reported that Meta employees asked to test the model for software development tasks have continued to prefer Anthropic's Claude. Wang has acknowledged the model trails rivals in coding, even as it draws praise for visual understanding. Some insiders, per the FT, have compared parts of the system to DeepSeek's latest model, while others note Muse Spark leans on Llama 4 code and datasets, despite Wang once describing it as built "from scratch. " Access has been narrow too. The model lives mostly inside Meta's own apps, with a private API rollout the FT describes as limited.

The ad machine still pays for the AI bill

The broader pressure on Wang is financial. The Wall Street Journal reported this week that 97.6% of Meta's 2025 revenue came from advertising, and that the company's planned AI capital spending this year is steeper relative to its size than Google's, Microsoft's, or Amazon's. Zuckerberg is now testing $4-a-month subscription tiers on Instagram, Facebook and WhatsApp, plus a $7.99 Meta AI chatbot subscription in select markets, hunting for revenue outside ads. Analysts at Truist Securities, cited by the WSJ, peg the subscription opportunity at as much as $20 billion annually by 2030, while Deutsche Bank floated $15.6 billion next year. Heady forecasts for a company that didn't clear $5 billion of non-ad revenue last year. Whether Wang's lab delivers the model to justify those numbers is the question Meta hasn't answered yet.