[Technology Comprehensive Report] On July 23rd, according to the latest disclosure by foreign media, there has been a fierce division within Apple over its open-source strategy for artificial intelligence (AI) technology. A team responsible for AI foundational model development had proposed earlier this year to open-source several proprietary models in order to attract external researchers to collaborate and optimize the technology, but this plan was ultimately rejected directly by Craig Federighi, the Senior Vice President of Software Engineering at Apple. This decision not only exposed deep contradictions within Apple’s AI strategy but also became one of the triggers leading to core team member departures.
Two insiders revealed that Pang Ruoming, the head of Apple’s AI foundational model team, had led the push for an open-source initiative. Internal documents show that the team even drafted a preliminary roadmap for open-sourcing, planning to prioritize the early version of the lightweight language model “Ajax LLM”—a model specifically optimized for mobile devices like iPhones, with parameters reduced by 90% compared to cloud-based models while speeding up inference by threefold.
However, Federighi explicitly opposed this plan in a reply email to Pang Ruoming. He believed that open-sourcing would force Apple to reveal details of its compression algorithms, allowing competitors to replicate its “performance for volume” edge optimization strategy, thereby undermining the AI differentiation advantage of iPhones.
More crucially, Federighi was concerned that open-sourcing would expose Apple’s fatal flaw: to adapt to the A series chips used in iPhones, its models must be compressed down to 1/10th their original size through techniques such as量化剪枝, which results in a more than 20% decrease in accuracy on complex tasks like multi-turn dialogue and long text generation compared to cloud versions. If external researchers develop more efficient compression schemes targeting this flaw, Apple might lose its lead in edge AI.
Federighi’s veto sparked strong dissatisfaction among the AI team. Several researchers revealed to the media that Apple’s “device-first” strategy has severely constrained technological development.
It is worth noting that Federighi himself acknowledged in a meeting with Pang Ruoming that “we may have gone too far on edge AI, but this is the strategy laid out by Cook and must be executed.” This statement was interpreted as a compromise and helplessness within Apple’s AI strategy.
According to supply chain news, Apple plans to launch the iPhone 18 series in 2026 with a new Siri 2.0, based on the upgraded Ajax LLM model. If this version still fails to narrow the performance gap with cloud models like GPT-5, Federighi’s “device-first” strategy might face greater scrutiny.
Meanwhile, there are reports that Apple is secretly preparing a “hybrid AI” project, attempting to balance performance and privacy through a hybrid architecture between edge and cloud computing—for example, assigning simple tasks to local processing on iPhones and complex tasks to upload to private cloud servers. If this plan can be implemented, it might provide a compromise path for the open-source controversy. (Qingshan)