Apple Silicon
The model is the silicon.
Available starting fall 2027
M·AI doesn't just run models — it becomes them. With model weights etched directly into read-only memory, inference happens at the speed of silicon.
Testing conducted by Apple in September 2027 using preproduction M·AI hardware. All comparisons are approximate and based on internal benchmarks.
A new kind of chip
For the first time, an AI model's weights are manufactured directly into silicon — not stored in memory, not loaded from disk. They exist as physical circuits, ready at the speed of electricity.
Architecture
M·AI reimagines the system-on-chip as a model-on-chip. The foundational model lives in read-only fabric, while a personal layer adapts in real time — learning your patterns without ever leaving the device.
The Secure Enclave+ ensures your personal fabric is cryptographically isolated, even from Apple.
Explore the architecturePersonal Fabric learns your patterns, preferences, and context — adapting the foundation model to you, entirely on device.
Your personal layer is just 42 MB of on-device LoRA parameters — expressive enough to deeply adapt, small enough to stay entirely on chip.
When the model lives in silicon, your data never needs to leave. Personal Fabric processes everything on device, in a cryptographically sealed enclave.
Input processed locally by Neural Engine
Personal Fabric encrypted at rest and in use
Immutable model weights in physical silicon
Only when needed. End-to-end encrypted.
Intelligence, evolved
Every interaction makes M·AI more useful — your writing style, your calendar patterns, your creative preferences. All learned locally. All encrypted. All yours.
Contextual Intelligence
The Intent Engine observes your context and surfaces what you need — generated entirely on device through your Personal Fabric.
One architecture, tuned for each form factor. Your Personal Fabric syncs seamlessly.
Your Personal Fabric transfers seamlessly between devices. Set up takes seconds — your intelligence follows you.
Use your Apple Account on any M·AI device.
Personal Fabric migrates via encrypted peer-to-peer.
The model resumes learning, instantly familiar.
Available fall 2027.
The Technology
Key innovations behind M·AI.
38.6 billion parameters manufactured directly into silicon using TSMC's N3E process. Model weights exist as physical circuit pathways — not bits in memory.
Low-rank adaptation layers that continuously learn your patterns. Just 42 MB of parameters, stored in the Secure Enclave+, adapting in real time.
A dedicated security processor that cryptographically isolates your Personal Fabric. Even Apple cannot access your learned patterns.
Proactive intelligence that observes your context — calendar, messages, location — and acts before you ask. All processing happens on device.
A dedicated neural processing unit delivering 42 trillion operations per second for inference, training, and real-time adaptation.
Your Personal Fabric syncs between M·AI devices using end-to-end encrypted peer-to-peer transfer. Your intelligence follows you everywhere.