'Apple talks to start-up about tech that runs large AI model on iPhone'

Apple is in discussions with PrismML about using the startup’s technology, The Information reports. PrismML claims it has gotten Qwen 3.6, a very large open source AI model from Alibaba, running on an iPhone 17 Pro.

The Information does not write exactly how advanced the discussions between Apple and PrismML are. The medium previously wrote that the iPhone manufacturer is investigating acquisition candidates that can help reduce AI models. This way, the models can run on devices instead of in the cloud.

PrismML shrinks language models by reducing the weights to 1bit or 1.58bit. The company claims to have reduced Qwen 3.6, a model with 27 billion active parameters, so that it can run on an iPhone 17 Pro. Apple’s AFM 3 Core Advanced model has 20 billion parameters, but uses a sparse architecture. As a result, only 1 to 4 billion parameters are active at a time.

PrismML introduced the Bonsai models at the end of March. Those are native 1bit models trained from scratch with 1bit weights, rather than 16bit or 32bit models compressed later. The Bonsai model with 8 billion parameters fits in 1.15GB of memory.