Meta’s Chief AI Officer Alexandr Wang admitted that the company’s open-source AI strategy no longer suits its frontier models, noting that the Muse Spark model remains proprietary due to safety risks identified during early training. In a June 14, 2026, interview on Bloomberg Tech, Wang stated that internal testing flagged bio risk and other elevated concerns that the company could not safely manage if the model weights were released to the public. He emphasized that this scaling challenge is currently affecting the entire AI industry as models become significantly more powerful.

Meta, which gained an industry reputation for its Llama family of open-weights models, is recalibrating its release framework. While the company intends to keep open-sourcing models it deems safe, it plans to restrict its most advanced frontier work. Wang is currently updating the company's internal advanced AI scaling framework to better evaluate and mitigate these risks. He noted that keeping Muse Spark proprietary allows Meta to maintain guardrails that are otherwise impossible to enforce once weights are widely accessible.

Despite these efforts, Muse Spark has faced performance criticism. Reports indicate that Meta employees testing the model for software development tasks prefer Anthropic’s Claude, as Muse Spark reportedly trails competitors in coding capabilities. Some internal feedback suggests the model relies heavily on Llama 4 code and datasets, contradicting initial descriptions of it being built from scratch. Access currently remains restricted primarily to internal applications with a limited private API rollout.

Financial pressure remains a central factor in Meta's strategy. According to the Wall Street Journal, 97.6% of Meta's 2025 revenue was derived from advertising, and the company’s planned AI capital expenditure currently exceeds that of Google, Microsoft, and Amazon relative to its size. To diversify revenue streams, Meta is testing subscription tiers across its platforms, including a $4-a-month subscription on Instagram, Facebook, and WhatsApp, and a $7.99 Meta AI chatbot subscription in select markets. Analysts at Truist Securities project these subscriptions could generate up to $20 billion annually by 2030, though the company has yet to clarify if its current AI models will successfully drive that growth.

Meta’s Chief AI Officer Alexandr Wang admitted that the company’s open-source AI strategy no longer suits its frontier models, noting that the Muse Spark model remains proprietary due to safety risks identified during early training. In a June 14, 2026, interview on Bloomberg Tech, Wang stated that internal testing flagged bio risk and other elevated concerns that the company could not safely manage if the model weights were released to the public. He emphasized that this scaling challenge is currently affecting the entire AI industry as models become significantly more powerful.

Meta, which gained an industry reputation for its Llama family of open-weights models, is recalibrating its release framework. While the company intends to keep open-sourcing models it deems safe, it plans to restrict its most advanced frontier work. Wang is currently updating the company's internal advanced AI scaling framework to better evaluate and mitigate these risks. He noted that keeping Muse Spark proprietary allows Meta to maintain guardrails that are otherwise impossible to enforce once weights are widely accessible.

Despite these efforts, Muse Spark has faced performance criticism. Reports indicate that Meta employees testing the model for software development tasks prefer Anthropic’s Claude, as Muse Spark reportedly trails competitors in coding capabilities. Some internal feedback suggests the model relies heavily on Llama 4 code and datasets, contradicting initial descriptions of it being built from scratch. Access currently remains restricted primarily to internal applications with a limited private API rollout.

Financial pressure remains a central factor in Meta's strategy. According to the Wall Street Journal, 97.6% of Meta's 2025 revenue was derived from advertising, and the company’s planned AI capital expenditure currently exceeds that of Google, Microsoft, and Amazon relative to its size. To diversify revenue streams, Meta is testing subscription tiers across its platforms, including a $4-a-month subscription on Instagram, Facebook, and WhatsApp, and a $7.99 Meta AI chatbot subscription in select markets. Analysts at Truist Securities project these subscriptions could generate up to $20 billion annually by 2030, though the company has yet to clarify if its current AI models will successfully drive that growth.