The trend toward small language models that can efficiently run on a single device instead of requiring cloud connectivity has emerged as a focus for Big Tech companies involved in artificial intelligence. Apple has released the OpenELM family of open-source models as its entry in that field. OpenELM uses “a layer-wise scaling strategy” to efficiently allocate parameters within each layer of the transformer model, resulting in what Apple claims is “enhanced accuracy.” The “ELM” stands for “Efficient Language Models,” and one media outlet couches it as “the future of AI on the iPhone.”

While OpenELM has yet to be tested publicly, an Apple Machine Learning Research post explains that at one billion parameters it “exhibits a 2.36 percent improvement in accuracy” with 2x fewer pre-training tokens when compared to the OLMo open source model from the Allen Institute.

Available on Hugging Face, “there are eight OpenELM models in total — four pre-trained and four instruction-tuned — covering different parameter sizes between 270 million and 3 billion parameters,” explains VentureBeat. Apple CoreNet instruction tools are available at GitHub.

Apple is making the OpenELM model weights available under what it says is a “sample code license,” writes VentureBeat, which notes that comes with “different checkpoints from training, stats on how the models perform as well as instructions for pre-training, evaluation, instruction tuning and parameter-efficient fine-tuning.”

At arXiv.org, an Apple research paper details OpenELM’s inference framework. “The sample code license does not prohibit commercial usage or modification,” VentureBeat says, calling the new family “the latest in a surprising string of open-source AI model releases from Apple, a notoriously secretive and typically ‘closed’ technology company, which has yet to publicly announce or discuss its efforts in this domain.”

In Q4, Apple quietly issued the open-source Ferret model, with multimodal capabilities. Ferret offers what CMSWire calls a “unique approach,” using spatial referencing for “superior accuracy in identifying elements within images, a novel feature compared to existing AI models.”

“OpenELM is a framework designed to work well on edge devices such as smartphones or laptops,” writes Tom’s Guide, which is “important for Apple as running AI locally is more secure and better for privacy.”

Tom’s points out “there is no indication of whether these models will form part of Apple’s plans for on-device AI in iOS 18 or upgrades to Siri, but they do show the direction the company is going with its AI.”