GitHub here: https://github.com/allenai/OLMoE.swift
Ai2 make as much of their training data available as possible: https://huggingface.co/datasets/allenai/OLMoE-mix-0924
I wonder how bit it is, normally they call them something with 3B or 7b
It’s 7 billion parameters big.
Ah, nice, not bad for a iPhone model.
Why only iPhone?
It probably uses Apple’s CoreML. Not sure if Android has an equivalent API.
Explain it pls
It’s Apple’s API for hardware-accelerated machine learning. Their new devices have specialized ML hardware so it’s actually feasible to run on mobile devices. Probably what they also use for Apple Intelligence. https://developer.apple.com/documentation/coreml
Seems like a good project to implement on the green side (Or pinguin side)
according to the linked article, it sounds like it is for performance reasons.
Wanna see how you git clone and compile it on ios
