Apple Has Launched OpenELM, Which Can Run Locally on Devices
Apple has launched OpenELM, a family of small, open-source AI models that can run directly on devices. These models, available on Hugging Face, offer efficient text generation and are optimized for both laptops and smartphones. OpenELM represents an exciting development in open-source
As Google, Samsung, and Microsoft continue their push for generative AI on PCs and mobile devices, Apple is now making its entrance with OpenELM, a new family of open-source language models (LLMs) that can run entirely on a single device without needing to connect to cloud servers.
Efficient AI Models for Text Generation
OpenELM, recently released on the AI code community Hugging Face, consists of small models designed to efficiently perform text generation tasks. There are eight OpenELM models in total – four pretrained and four instruction-tuned – with parameter counts ranging between 270 million and 3 billion.
Pretraining and Instruction Tuning
Pretraining is a method for getting a language model to produce coherent text, but it's primarily a predictive exercise. Instruction tuning, on the other hand, makes the model better at responding with relevant output to specific requests. For example, a pretrained model might respond to "teach me to bake bread" with "in a home oven," while an instruction-tuned model would give detailed step-by-step instructions.
Open Source With Commercial Possibilities
Apple offers the OpenELM models' weights under a "sample code license," along with various training checkpoints, performance statistics, and instructions for pretraining, evaluation, instruction tuning, and parameter-efficient fine-tuning. The license permits commercial use and modification, with requirements to retain certain information when redistributing the code.
Safety and Responsibility
Apple notes that the models "are provided without any safety guarantees," meaning they may generate inaccurate, harmful, biased, or offensive responses. This is part of a surprising series of open-source AI model releases from Apple, which is usually a very closed-off tech giant.
Technical Specifications and Performance
The OpenELM models, ranging from 270 million to 3 billion parameters, were pretrained on public datasets containing 1.8 trillion tokens from sources such as Reddit, Wikipedia, and arXiv.org. The models are optimized to run on ordinary laptops or even certain smartphones.
Apple uses a layer-wise scaling strategy to distribute parameters within each layer of the transformer model, improving accuracy while using computational resources efficiently. Performance-wise, Apple's results show the OpenELM models perform respectably, with particularly strong results from the 450 million parameter instruction-tuned variant.
Future Development and Potential
Although OpenELM hasn't yet been widely tested, Apple's publication on Hugging Face shows it's aiming for applications that can run directly on devices, similar to Google's, Samsung's, and Microsoft's latest models. Over time, OpenELM is expected to improve, and it will be interesting to see how the community makes use of Apple's open-source initiative across various applications.