Skip to content

Built-in Mac Model

macOS 26 and later include Apple's own on-device language model, the one Apple Intelligence uses through the FoundationModels framework. AI:GO serves it locally through the bundled aigo-apple-llm-server sidecar on both the desktop app and headless aigo-server (exposing an OpenAI-compatible inference API, with readiness queryable at GET /api/v1/models/apple-on-device/status), so a fresh installation can chat, title conversations, and summarize text before you download anything. Any orphaned aigo-apple-llm-server process left behind after an unexpected exit is cleaned up automatically on the next launch. Because the model and its weights belong to macOS, there is nothing to download and nothing to delete in AI:GO.

Which Macs can use it

The model runs on a Mac with Apple Silicon and macOS 26 or later, with Apple Intelligence turned on in System Settings > Apple Intelligence & Siri and its model downloaded by macOS.

On macOS 15, on Macs that cannot run Apple Intelligence, and on Linux and Windows, AI:GO shows nothing about it. On a Mac that supports it while Apple Intelligence is turned off or still downloading the model, AI:GO lists the entry along with a one-step fix (see When it cannot run).

Its name

AI:GO shows the model under the name Apple gives it, identical across UI languages. On macOS 27 that is the variant your Mac runs, AFM 3 Core or AFM 3 Core Advanced (macOS chooses the variant automatically). On macOS 26 the name is Apple Foundation Model. Beside the name, a Built into Mac tag and the Apple logo mark it as part of macOS.

Where it appears

In Models > Manage, it appears as the first row under This computer, with its context window tag and the standard Run and Stop actions. Its ⋯ menu offers Chat with this model; it has no settings dialog, no file size, and no delete action. On a fresh installation where it is the only local model, Manage displays it directly instead of the empty-state download prompt.

In the chat model picker, it sits right below currently running models, and selecting it starts the server automatically. In Settings > Models, while the model can answer, you can assign it as Chat under Default models, or under Models for tasks as Titles and summaries (used for conversation titles) or Data wiki (used for summaries). Agent and Squad model pickers also list it, though it is never chosen automatically as an agent's default.

The first chat model

When no model is chosen and nothing has been downloaded yet, chat starts on the built-in Mac model without writing a permanent preference, so downloading or choosing another model takes over cleanly.

  • A download: once your first downloaded language model finishes, it becomes the active chat model, even if you had clicked the built-in model in a single-item picker earlier.
  • A default in Settings: if you explicitly set the built-in model as the chat default in Settings > Models > Default models, finishing a download does not replace your choice.
  • A pick in the chat picker: selecting any model in the chat picker after at least one download keeps that selection, including selecting the built-in model again.

The built-in model remains in your model list after you download other models, ready whenever you want a zero-VRAM local reply.

When it cannot run

When the model is listed on a supported Mac yet cannot answer right now, its row in Models > Manage states the reason on one line and replaces Run with the action that resolves it. The chat picker shows the same message and triggers the same action when clicked.

What the row says Button What to do
Apple Intelligence is turned off. Open System Settings Turn on Apple Intelligence in Apple Intelligence & Siri, then return to AI:GO.
Apple Intelligence is still downloading the model. Check again Wait for macOS to finish downloading the model.
Could not check its status. Retry Check again; if it keeps failing, restart AI:GO.
It does not support the system language. Language & Region link Choose a supported language in System Settings > General > Language & Region. The model still runs.

AI:GO re-checks status automatically whenever its window regains focus while one of these guidance lines is visible, so turning Apple Intelligence on takes effect without restarting the app. If an availability check times out transiently, AI:GO preserves the model's known name in the picker instead of hiding it, and a failed start re-checks availability once to surface updated guidance. In a browser connected to aigo-server, the row names the System Settings pane rather than opening it directly.

What it is good for

The built-in model handles quick chat, conversation titles, and short summaries. Its context window is queried from macOS at runtime (8,192 tokens measured on macOS 27.2, and 4,096 tokens on macOS 26), and Models > Manage displays the exact token limit your Mac reports.

Because the on-device Apple model does not execute multi-step tool loops or accept image inputs in this path, agents, long-context document analysis, and vision questions should use a downloaded GGUF or MLX model from Models > Explore. General chat applies Apple's standard safety guardrails, while title and summary generation on user-supplied text uses Apple's permissive content-transformation mode.