Apple's Foundation Models — an on-device LLM in a few lines of Swift 🧠

At WWDC 2025 Apple did the thing everyone wanted: it opened the on-device large language model behind Apple Intelligence to developers, through the new Foundation Models framework. Ships with macOS Tahoe (26) beta and Xcode 26, runs locally — so the usual privacy story, no API key, no cloud cost, no network needed.

The part that surprised me is how small the API is. Here’s a working session, start to finish:

let session = LanguageModelSession(instructions: """
    You are a motivational workout coach that provides quotes to inspire \
    and motivate athletes.
    """
)

let prompt = "Generate a motivational quote for my next workout."
let response = try await session.respond(to: prompt)
print("\(response.content)")

That’s it. An instructions string, a prompt, an await. No model loading, no tokenizer wrangling, no server. For something that’s a full LLM running on the machine, the ergonomics are hard to beat.

My experiments

I poked at it from a couple of angles:

FoundationModelCli — a small command-line app, just to confirm the thing works outside an app target and to feel out the API. (Spoiler: see the snippet above — it really is that simple.)
mcp-foundation-models — an MCP server that wraps the framework, so any MCP client can reach Apple’s local model. That’s the one I find most interesting: a private, on-device model exposed over a standard protocol that the wider agent ecosystem already speaks.
foundation-models-plugin — a 4D plugin giving 4D code a direct path to the on-device model. Rougher than the other two, but it closes the loop: Swift, the MCP ecosystem, and 4D can now all reach the same local LLM.

Why this matters

The framework itself is the headline, but the second-order effect is the fun one. Once a private on-device model has a clean API, you can put it behind whatever transport you like — a CLI, an HTTP server, an MCP server — and suddenly your existing tools can use a model that never phones home. Small API, big surface area.

Worth reading

Apple shipping a usable, on-device LLM API with no key and no bill is a quietly big deal — and the moment it exists, wrapping it in a CLI or an MCP server is almost irresistible. 🧠

EDIT — and a year later, Apple shipped its `fm`

Looking back: putting the on-device model behind a CLI, an MCP server, and a 4D plugin turned out to be the whole story — “small API, big surface area” held up. The punchline came a year on, when macOS 27 shipped a command-line tool called fm… the same name I’d given mine. The full tale is here. 🧠

Apple's Foundation Models — an on-device LLM in a few lines of Swift 🧠

My experiments

Why this matters

Worth reading

EDIT — and a year later, Apple shipped its fm

EDIT — and a year later, Apple shipped its `fm`