The Apple’s AI models that underpin Apple Intelligence’s capabilities on iOS, macOS, and other platforms have been updated, according to Apple. These new models, however, now perform worse than those of top rivals like OpenAI, Google, and Meta, based on the company’s benchmarks.
Human reviewers evaluated the text quality produced by Apple’s most recent AI model “Apple On-Device” as being on pace with, but not superior to, similarly sized models from Google and Alibaba, the company said in a blog post published Monday.
About 3 billion characteristics make up the Apple On-Device model, which runs offline on gadgets like the iPhone and offers functions like text analysis and summarization.
Additionally, Apple unveiled “Apple Server,” a more sophisticated variant intended to operate in the business’s data centers. Human testers evaluated Apple Server’s text generating capabilities worse than OpenAI’s GPT-4o, a model that has been on the market for more than a year, based on Apple’s benchmarks.
Apple assessed its models’ capacity to analyze visuals in addition to producing text. For image processing tasks, human raters favored Meta’s Llama 4 Scout model over Apple Server.
The fact that Llama 4 Scout itself does not constantly perform better than top models from Google, Anthropic, and OpenAI underscores the difficulties Apple has in this market.
Apple said that both the Apple On-Device as well as Apple Server models provide better tool use and productivity than the previous versions, in spite of these benchmark results.
About 15 languages can be comprehended by the models, due to an extended training dataset that includes photos, PDFs, texts, infographics, tables, and charts. This week, the company’s Foundation Models framework opened up the Apple On-Device concept to outside developers.
Loading ...
