Apple Foundation Models

Apple Foundation Models is the on-device language model Apple ships inside iOS 26/macOS 26: ~3B parameters quantized to 2 bits (~1GB footprint), with a 4K token context window (doubled to 8K in the iOS 27 generation). Any app can call it through the FoundationModels Swift framework — summarization, extraction, classification, entity detection, short-form generation — with zero network calls, zero per-request cost, and data that never leaves the device. Apple also exposes a larger Private Cloud Compute model (32K context) for work the on-device model can't handle, under the same no-storage privacy posture.

Be clear-eyed about what it isn't: this is a small utility model, not a general assistant — Apple itself says it is not suited for world knowledge, complex reasoning, math, or translation. The weights are not publicly downloadable (they ship with the OS), so openWeights is false in the downloadable sense. But for developers in the Apple ecosystem, it is the default private-AI substrate: free inference, on by default, on hundreds of millions of devices.