Apple’s upgraded AI models underwhelm on performance

Apple has announced updates on AI models that strengthen the suit of Apple intelligence features throughout iOS, Macos and many more. However, according to the company’s own criteria, models reduce old models from rival technology companies, including OpenAI.

Apple Says in a blog post Human examiners who have raised the quality of the text produced by its new “Apple on-Devis” model on Monday-which are run offline on products, including iPhones-“relatively” is better than “the same size” better than the texts of Alibaba models. Meanwhile, those same testers have rated Apple’s new model, called “Apple Server” and designed to run the OpenAI -old back company data center. GPT -4OThe

In a separate experiment evaluating the ability to analyze the images of Apple models, human rates according to Apple prefer the meta Lama 4 scout model than the Apple server, according to Apple. It is somewhat surprised. In several exams, Call 4 Scouts AI labs such as Google, ethnographic and Open are acting worse than the top models from the labs.

Benchmark results in the report that Apple’s AI Research Department suggests to fight to contestants to hold competitors in the Cutthrote AI race. Apple has AI capacity in recent years UnderhalemedAnd a promised Seri upgrade has been delayed indefinitely. Some customer The apple has suedThe farm complains of marketing AI features for its products that it has not yet provided.

In addition to generating the text, Apple drives on-devices, which in size about 3 billion parameters, shortage and textual analysis drives. (Parameters are roughly matched to solve a model of a model, and more parameter models usually perform better than low parameters)) Third -party developers can tap it through Apple’s Foundation Model Framework until Monday.

Apple says that both Apple on device and Apple server are proud of the use and skills of advanced equipment compared to their predecessors and understand about 15 languages. This is thanks to the part of an extended training dataset that includes image data, PDFS, documents, manuscripts, infographics, tables and charts.