# Apple Foundation Models on-device diagnosis as an alternative to a cloud AI call > Source-anchored research for Apple Foundation Models on-device diagnosis as an alternative to a cloud AI call. Run status: COMPLETED. Parent: [Apple Platform Frontier for Mac Utilities](https://llms-explorer.com/tree/apple-platform-frontier-for-mac-utilities/) · 5 facets · 10 facts · page: https://llms-explorer.com/tree/apple-foundation-models-on-device-diagnosis-as-an-alternative-to-a-cloud-ai-call/ ## How it works - Availability is not guaranteed: SystemLanguageModel.default.availability returns .available or .unavailable with reasons deviceNotEligible, appleIntelligenceNotEnabled or modelNotReady, and Apple tells developers to check it before creating a LanguageModelSession and show a fallback. A Mac App Store diagnosis feature must therefore degrade to non-AI output on ineligible Macs or when the user has Apple Intelligence off. (high confidence) — [source](https://developer.apple.com/events/resources/code-along-205/) - WWDC26 introduces a LanguageModel protocol so session code is unchanged when swapping the Apple model for a third-party one (Anthropic and Google publish Swift packages; open-source CoreAILanguageModel and MLXLanguageModel run local models). A monitor can develop against SystemLanguageModel and make cloud a user-chosen option. (medium confidence) (verified as of 2026-06-08) — [source](https://developer.apple.com/videos/play/wwdc2026/241/) ## Measurements and reference values - Apple's own evaluation shows the on-device model beating Qwen-2.5-3B across languages and competitive with Qwen-3-4B and Gemma-3-4B in English, with a ~4.6% MGSM regression from compression; quality is mid-size-open-model level, below cloud frontier models for reasoning over large contexts. (medium confidence) — [source](https://machinelearning.apple.com/research/apple-foundation-models-2025-updates) ## Problems, failure modes and limitations - An independent arXiv audit (24 Aug 2026) of SystemLanguageModel.default found it confabulated confidently on 69.1% of false-premise questions, refused 18.4% of benign summarization inputs (peers 0%), and had verbalized confidence near 99% with chance-level AUROC 0.47. A diagnosis feature should not present model output as authoritative; ground it in measured data and let deterministic code decide pass/fail. (low confidence) (verified as of 2026-08-24) — [source](https://arxiv.org/html/2608.23663v1) ## Comparisons and alternatives - WWDC26 adds PrivateCloudComputeLanguageModel behind the same session API: 32,000-token context, reasoning levels (.light/.deep), no account or API key, prompts not stored, free for developers under 2 million first-time downloads with higher daily limits for iCloud+ users. It is a cloud call, so it needs the network and its privacy claim is Apple's PCC guarantee rather than on-device. (medium confidence) (verified as of 2026-06-08) — [source](https://developer.apple.com/videos/play/wwdc2026/241/) ## Facts and statements - The Foundation Models framework exposes Apple's roughly 3-billion-parameter on-device model to apps through a Swift API; it runs offline and Apple states it is designed for summarization, entity extraction, text understanding, refinement and short dialog, not as a general world-knowledge chatbot. (high confidence) — [source](https://machinelearning.apple.com/research/apple-foundation-models-2025-updates) - The WWDC26 release ships a rebuilt on-device model, better at logic and tool calling and with vision input; the session example prints model.contextSize as 8192 and adds model.tokenCount(for:) so an app can budget instructions and prompts. (medium confidence) (verified as of 2026-06-08) — [source](https://developer.apple.com/videos/play/wwdc2026/241/) - The original (2025) on-device model had a fixed 4096-token context window shared by instructions, prompt, tool schemas and response; exceeding it throws GenerationError.exceededContextWindowSize. A raw DNS or connection log therefore has to be pre-aggregated in app code before it reaches the model. (medium confidence) (verified as of 2026-06-08) — [source](https://developer.apple.com/forums/thread/806542) - On-device inference keeps prompt and response data on the Mac, so hostnames, IPs and resolver config in a diagnosis prompt are not sent to a third party; this is the privacy advantage over a cloud AI call. Apple also states it does not train on users' private personal data or interactions. (high confidence) — [source](https://machinelearning.apple.com/research/apple-foundation-models-2025-updates) - WWDC26 adds an fm command-line tool (macOS 27), a Python SDK, an Evaluations framework to measure feature quality, and open-sources the framework; the Evaluations framework is the sanctioned way to test diagnosis prompt quality before shipping. (low confidence) (verified as of 2026-06-08) — [source](https://developer.apple.com/videos/play/wwdc2026/241/)