Apple Foundation Models
An on-device foundation model framework introduced by Apple, providing iOS applications with a free, offline 3-billion-parameter language model. It supports tool invocation via the Tool protocol and incorporates privacy-preserving characteristics.
Timeline (last 90 days)
Do not trust availability flags self-reported by engines; make one real single-token probe call and cache the result.
Apple Foundation Models require no download and come bundled with the system; the time-to-first-token for the first call is 9.7–19.2 seconds, the median time-to-first-token for subsequent calls is 580 ms, and the model runs in a system process without consuming application memory.
Apple's test figures were obtained on a Mac (M1/8 GB) and cannot be used to predict latency on real devices; they can only be used for a side-by-side comparison between the two engines.
In the iOS 26.3 simulator, apple.isAvailable() returns true but all 30 calls fail with GenerationError error -1, because the model directory contains no assets; the same model works correctly in a Swift environment on macOS 26.6.
On a structured output task involving a store audit checklist, ExecuTorch (Llama 3.2 1B, schema stuffed into the prompt) produced only 2 valid outputs out of 20 calls, while Apple Foundation Models (guided generation) produced valid outputs for all 20 calls.
Apple Foundation Models constrains model output at the framework level via guided generation to necessarily conform to a given type structure, whereas ExecuTorch has no corresponding capability and can only write the schema into the prompt.
Each session in Apple Foundation Models has a context window of 4,096 tokens.
The Tool protocol in the Apple Foundation Models framework allows developers to register external functions as tools that the model can call; the model decides when to call which tool and what parameters to pass, while the actual execution logic is completed by developer-controlled Swift code.
Apple's Foundation Models framework provides iOS applications with a free, offline 3-billion-parameter model.
All Facts (9)
Do not trust availability flags self-reported by engines; make one real single-token probe call and cache the result.
50%UnverifiedApple Foundation Models require no download and come bundled with the system; the time-to-first-token for the first call is 9.7–19.2 seconds, the median time-to-first-token for subsequent calls is 580 ms, and the model runs in a system process without consuming application memory.
50%UnverifiedApple's test figures were obtained on a Mac (M1/8 GB) and cannot be used to predict latency on real devices; they can only be used for a side-by-side comparison between the two engines.
50%UnverifiedIn the iOS 26.3 simulator, apple.isAvailable() returns true but all 30 calls fail with GenerationError error -1, because the model directory contains no assets; the same model works correctly in a Swift environment on macOS 26.6.
50%UnverifiedOn a structured output task involving a store audit checklist, ExecuTorch (Llama 3.2 1B, schema stuffed into the prompt) produced only 2 valid outputs out of 20 calls, while Apple Foundation Models (guided generation) produced valid outputs for all 20 calls.
50%UnverifiedApple Foundation Models constrains model output at the framework level via guided generation to necessarily conform to a given type structure, whereas ExecuTorch has no corresponding capability and can only write the schema into the prompt.
50%UnverifiedEach session in Apple Foundation Models has a context window of 4,096 tokens.
50%UnverifiedThe Tool protocol in the Apple Foundation Models framework allows developers to register external functions as tools that the model can call; the model decides when to call which tool and what parameters to pass, while the actual execution logic is completed by developer-controlled Swift code.
50%UnverifiedApple's Foundation Models framework provides iOS applications with a free, offline 3-billion-parameter model.
50%