Apple Picks Google Gemini to Power Its Next Foundation Models and Siri

Why It Matters for Developers#

The strategic lesson applies well beyond Apple. If a company with Apple's resources concluded that the best foundation for its models came from a specialist, most product teams should treat model choice as a sourcing decision, re-evaluated regularly against their own tasks. The AI evaluation guide shows how to build the kind of repeatable comparison that Apple described in general terms.

The architecture is the second lesson. Apple's design separates the model supplier from the execution environment and keeps sensitive processing on hardware it controls. .NET teams can apply the same idea at a smaller scale: send prompts that contain personal or regulated data to a model you host, and allow a cloud model for everything else. With Microsoft.Extensions.AI, both are just IChatClient registrations, and a small router decides per request:

using Microsoft.Extensions.AI;
using OllamaSharp;
using OpenAI.Chat;

var builder = WebApplication.CreateBuilder(args);

string Required(string key) =>
    builder.Configuration[key] ?? throw new InvalidOperationException($"Missing setting '{key}'.");

// Sensitive prompts go to a self-hosted model; other prompts may use a cloud model.
builder.Services.AddKeyedChatClient("local",
    new OllamaApiClient(new Uri(Required("Local:Endpoint")), Required("Local:Model")));
builder.Services.AddKeyedChatClient("cloud",
    new ChatClient(Required("Cloud:Model"), Required("Cloud:ApiKey")).AsIChatClient());
builder.Services.AddSingleton<AssistantRouter>();

var app = builder.Build();

app.MapPost("/assist", async (AssistantRouter router, AssistRequest req, CancellationToken ct) =>
    Results.Ok(await router.AskAsync(req.Prompt, req.ContainsPersonalData, ct)));

app.Run();

public sealed class AssistantRouter(
    [FromKeyedServices("local")] IChatClient local,
    [FromKeyedServices("cloud")] IChatClient cloud)
{
    public async Task<string> AskAsync(
        string prompt, bool containsPersonalData, CancellationToken ct)
    {
        var client = containsPersonalData ? local : cloud;
        var response = await client.GetResponseAsync(prompt, cancellationToken: ct);
        return response.Text;
    }
}

public sealed record AssistRequest(string Prompt, bool ContainsPersonalData);

In production, decide sensitivity on the server with your own data classification rather than trusting a client flag, and log which route served each request. The local AI guide covers running models with Ollama and ONNX Runtime, and the responsible AI guide covers the privacy and safety controls that should sit around both routes.

For developers who ship on Apple platforms, including with .NET MAUI, the practical effect will depend on what Apple exposes to apps once the new models arrive. The joint statement did not mention developer APIs, so any change to how third-party apps interact with Siri or Apple Intelligence remains to be announced.