Competition within the AI landscape has begun to pivot from the pursuit of building the largest models to crafting smaller yet highly capable tools. Recognizing the value in this shift, Apple has strategically positioned itself in the competitive artificial intelligence market by introducing eight compact AI models collectively known as OpenELM. These models are tailored to operate efficiently on devices and in offline settings, making them particularly suited for smartphone applications.

Published on the open-source AI platform Hugging Face, the OpenELM models are available in various sizes, ranging from 270 million to 3 billion parameters. Apple offers both pre-trained and instruction-tuned versions of these models. The pre-trained models serve as a foundation for users to fine-tune and customize, while the instruction-tuned variants are optimized for seamless interactions and conversations with end-users.

Although Apple has not explicitly outlined specific use cases for these models, they hold potential applications in tasks such as email and text parsing, as well as providing personalized suggestions based on user data. This strategy mirrors Google's deployment of its Gemini AI model across its Pixel smartphone lineup.

The models have been trained on publicly available datasets, and Apple has made the code for CoreNet (the training library for OpenELM) and the model "recipes" accessible to users. This transparency allows users to scrutinize the model-building process.

Apple's release of OpenELM comes shortly after Microsoft unveiled Phi-3, a family of small language models designed for local execution. Notably, Phi-3 Mini, with its 3.8 billion parameters and impressive token capacity, has demonstrated performance comparable to GPT-4, outperforming competitors such as Llama-3 and Mistral Large.

While Apple has yet to integrate these new AI language model capabilities into its consumer devices, rumors suggest that the upcoming iOS 18 update will introduce AI features leveraging on-device processing to enhance user privacy.

Apple's hardware possesses an edge in local AI utilization, as it seamlessly integrates device RAM with GPU video RAM, enabling efficient execution of AI models. In contrast, Windows devices face limitations due to separate device RAM and GPU VRAM, necessitating additional hardware configurations to run AI models effectively.

However, Apple trails behind Windows/Linux in AI development, predominantly reliant on hardware from Nvidia, which Apple phased out in favor of its proprietary chips. Consequently, there is a relative scarcity of Apple-native AI development, necessitating translation layers or complex procedures for AI utilization on Apple products.