Apple has released OpenELM, a family of open source generative AI language models. The apple brand wanted it to be efficient and less energy intensive. Soon on iOS 18 and macOS 15?
In turn, Apple this week presented a brand new family of language models (LLM), called “OpenELM”, which is intended for generative artificial intelligence. Usually very secretive, the American giant has chosen an open source strategy, to promote models tinged with sobriety, which should focus on specific uses and specific devices. They could be aimed at iPhones and others devices of the brand. Explanations.
Small language models intended to work locally, on Apple devices
At almost the same time as its illustrious competitor Microsoft, which introduced a smaller and more economical AI model, Phi-3, Apple unveiled OpenELM, a family of language models that its designers themselves describe as “ efficient and open source “.
The choice of open source, Meta did the same, but at Apple, it necessarily surprises a little. The Cupertino company has, however, published four versions of the model on Hugging Face and code snippets on GitHub. There are pre-trained models with 270 million, 450 million, 1.1 billion and 3 billion parameters, even smaller than the smallest of Microsoft’s Phi-3 models (3.8 billion parameters), presented in same time.
On the scale of the sometimes disproportionate LLMs of the giants of the sector, these language models made in Apple are… small, yes. But that doesn’t mean they aren’t effective or even less useful. Because if the most imposing models work on remote servers, Apple wants to run its AI locally, on its devices directly. And that’s the subtlety, in addition to significant energy savings.
An Apple generative AI, for what uses and why?
From what we can read on Hugging Face, we understand that the language models of the OpenELM family will be able to concentrate on targeted tasks, such as writing emails for example, by focusing on pure user uses. Which obviously makes sense, at a time when the French are increasingly using generative AI (+60% usage compared to 2023, according to the latest Ifop survey for Talan published on April 26, 2024).
The release schedule for these language models is not trivial. Apple is indeed preparing its WWDC (June 10), its annual conference during which the company will present the next updates of its systems, iOS 18 and macOS 15 to name but a few. If the Californian company has not hidden its ambition to “ run AI locally on Apple devices “, it is still difficult to know today when and how OpenELM will appear in the iPhone and other devices of the brand.
Developers can in any case already use them, while keeping in mind, as Apple indicates, that they are made available “ without any security guarantee ”, which can lead to “ inaccurate, harmful, biased or objectionable results ”, according to user requests (prompts). Apple has published the CoreNet library to help them train standard and new generative AI models.
Sources: Hugging Face, GitHub, Apple
0