Open-source generative AI appears to be developing rapidly, especially over the past week with several tech giants unveiling powerful AI models one after the other. French AI startup Mistral just unveiled the Mixtral 8x22B and Meta announced Llama 3 only a day later. Now Apple is here with its own announcement, albeit a bit different from its competitors.

Unlike Mistral and Meta’s cloud-based offerings, Apple’s new OpenELM of open-source AI models champion on-device AI for devices such as iPhones and MacBooks. This is because OpenELM features compact and efficient AI models that can easily be handled by consumer devices.

OpenELM is a family of not just two or three, but eight models with four different parameter sizes (270M, 450M, 1.1B, and 3B), all trained on public datasets, at least according to Apple.

As mentioned earlier, these are designed to run efficiently on devices, which brings the advantage of lower latency and added privacy compared to cloud-based counterparts since no data needs to be shared to the cloud over an internet connection.

The actual performance of these models, however, remains to be seen. But as far as benchmark scores go, OpenELM is marginally ahead of its rival open-source compact models like OLMo despite requiring 2x less training data.

Apple has open-sourced not only OpenELM, but also the library used to train the AI model called CoreNet. Additionally, models allowing for “efficient inference and fine-tuning on Apple devices” have been made open-source as well.

It appears that the trend of developing smaller models is becoming widespread, and this new release from Apple provides further indication of its forthcoming on-device AI initiatives expected at WWDC. Equally significant is the decision to make this release open-source, marking a clear departure from Apple’s historically guarded and secretive practices.

OpenELM directly rivals Microsoft’s recent Phi-3 AI model, which is similarly compact and open-source and even meant to run locally on devices such as smartphones and laptops. Despite being a compact AI model, Phi-3 is able to beat much larger competitors such as OpenAI’s free GPT 3.5 and Meta’s new Llama 3.