Very large AI models on the iPhone: Apple is interested in new technology from a startup | News

Large language models, so-called Large Language Models (LLM) such as ChatGPT, Microsoft Copilot or Google Gemini, are based on huge amounts of data. They are therefore mostly processed in data centers. Apple, on the other hand, is relying at least partially on local execution for Siri AI, the future AI-based generation of its own voice assistant. This should even be possible on devices such as iPhones and iPads, which only have limited RAM and SSD storage. To achieve this, reduced LLMs are used on smartphones and tablets from Cupertino; More complex user requests are forwarded to a corresponding cloud service. The Wall Street Journal recently tested how this works in practice using the beta version of iOS 27.

LLM storage requirements shrunk to less than ten percent
Being able to run features based on artificial intelligence locally presents manufacturers like Apple with the challenge of reducing the size of the corresponding models. This is usually accompanied by a restriction of options because only a smaller number of parameters are available. However, a startup called PrismML says it has managed to drastically reduce the memory requirements of large language models without reducing their size. According to a report by The Information (Paywall), the company says it has shrunk the open source LLM Qwen 3.6 developed by Alibaba from 54 gigabytes to just 4 gigabytes. This did not limit the performance of the model, PrismML told the news portal, as it still had over 27 billion active parameters even after compression. These values ​​roughly correspond to the number of synapses in the human brain.

27 billion simultaneously active parameters possible on the iPhone
With a memory requirement in the single-digit gigabyte range, the AI ​​model shrunk by PrismML is suitable for local use on devices such as the iPhone. This apparently didn’t go unnoticed in Cupertino: According to The Information, Apple has already contacted PrismML. Representatives of both companies have met several times to explore collaboration options. Apple is itself working on reducing the LLM size, but in the case of Siri AI it can only access 1 to 4 billion simultaneously active parameters, although the entire model includes around 20 billion parameters. This limitation does not exist with PrismML’s new technology, the report says. This makes them extremely interesting for the iPhone manufacturer. However, it is not yet clear whether Apple and PrismML will work together and what this might look like.