Official documentation

Deploying Transformers on the Apple Neural Engine

Apple Machine Learning Research, Apple Machine Learning Research, 2022

Read on October 16, 2026

What it is: Apple's June 2022 research article on running Transformer neural networks on the Apple Neural Engine in Apple silicon. It describes machine learning inference on the device and its privacy benefit.
How much weight it carries: Apple's own engineering write-up, with the reference code Apple released. It explains how Apple's hardware runs neural networks in general. It does not describe the Dictation model, so it supports the on-device principle, not a specific Dictation claim.

Facts taken from this source

Apple says the fifth-generation 16-core Neural Engine, reached by 2021, is capable of 15.8 teraflops, which it puts at 26 times the processing power of the original.
“the fifth-generation of the 16-core ANE is capable of 26 times the processing power, or 15.8 TFlops, of the original” (The Apple Neural Engine section)
Apple describes the Apple Neural Engine (ANE) as the energy-efficient and high-throughput engine for machine learning inference on Apple silicon.
“the energy-efficient and high-throughput engine for ML inference on Apple silicon” (Introduction)
Apple says that on-device machine learning benefits privacy because the data for inference workloads stays on the device rather than on a server.
“Increasing the adoption of on-device ML deployment will also benefit user privacy, since data for inference workloads remains on-device, not on the server.” (Introduction, Apple Neural Engine section)

Official documentation

Apple Machine Learning Research, Apple Machine Learning Research, 2022

Read on October 16, 2026

What it is: Apple's June 2022 research article on running Transformer neural networks on the Apple Neural Engine in Apple silicon. It describes machine learning inference on the device and its privacy benefit.
How much weight it carries: Apple's own engineering write-up, with the reference code Apple released. It explains how Apple's hardware runs neural networks in general. It does not describe the Dictation model, so it supports the on-device principle, not a specific Dictation claim.

Apple says the fifth-generation 16-core Neural Engine, reached by 2021, is capable of 15.8 teraflops, which it puts at 26 times the processing power of the original.
“the fifth-generation of the 16-core ANE is capable of 26 times the processing power, or 15.8 TFlops, of the original” (The Apple Neural Engine section)
Apple describes the Apple Neural Engine (ANE) as the energy-efficient and high-throughput engine for machine learning inference on Apple silicon.
“the energy-efficient and high-throughput engine for ML inference on Apple silicon” (Introduction)
Apple says that on-device machine learning benefits privacy because the data for inference workloads stays on the device rather than on a server.
“Increasing the adoption of on-device ML deployment will also benefit user privacy, since data for inference workloads remains on-device, not on the server.” (Introduction, Apple Neural Engine section)