WWDC26 Apple Intelligence: A Developer’s Deep Dive After 3 Months

Apple Intelligence shipped with macOS 15.4. After 3 months of daily use, here’s the honest deep dive — what’s good, what’s broken, and what matters for developers.

What Shipped (the actual list)

Writing Tools — Rewrite, Proofread, Summarize (system-wide)
Image Playground — text-to-image (on-device, 3B param model)
Genmoji — text-to-emoji (on-device, similar architecture)
Siri 2.0 — LLM-powered, on-device for simple queries, Private Cloud Compute for complex
Image Wand — turn rough sketches into polished images
Priority Messages — LLM-summarized emails
Smart Reply — context-aware quick responses

What Works (3-month test)

Writing Tools

Proofread: catches 90% of typos, including context-specific ones
Rewrite: 4 styles (Professional, Friendly, Concise, Standard) — useful
Summarize: works well for long emails, less so for technical docs

Siri 2.0

Voice commands now feel natural (3B on-device model)
Cross-app actions (e.g., “send this photo to John in Messages”) work reliably
“When did Mom’s flight land?” type questions are actually answered correctly

What’s Broken

Image Playground

Output quality is mediocre. 512x512 generation, often distorted
Style consistency is poor (asks for “a cat in watercolor” gets 5 different cat styles)
Apple says it improves over time — it does, slowly

Genmoji

Cute but limited — only 4-5 styles per emoji
For “make me a happy cat” it gives a yellow cat, not much variety

Siri Web Search

When you ask “find me the best Chinese restaurant nearby” it often returns random Google results, not contextual
Apple promised Private Cloud Compute for privacy — but it requires internet, so the privacy claim is dubious

What Matters for Developers

Foundation Model API

Apple opened up the Foundation Model API in macOS 15.4. You can now call the on-device LLM from your apps:

import FoundationModels

let model = SystemLanguageModel.default
let session = LanguageModelSession(model: model)

let response = try await session.respond(
    to: "Explain quantum entanglement in 3 sentences"
)
print(response.content)

Privacy by design — runs locally, no network. Useful for:

Email classification
Smart replies
Content moderation
Code autocomplete (yes, it does this well)

Xcode AI Integration

Xcode 16 added AI-powered features:

Smart code completion (uses Foundation Model, not OpenAI)
Doc generation from code
“Explain this function” hover docs

I tested these — they’re useful but not transformative. Similar to Copilot quality.

My Verdict

For users: Apple Intelligence is solid but not transformative. 70% of features are useful (Writing Tools, Siri 2.0), 30% are mediocre (Image Playground, Genmoji).

For developers: Foundation Model API is a big deal. Free, fast, private on-device LLM access. Use it for any user-facing app that needs natural language understanding.

For the industry: Apple showed that on-device LLMs are viable for production. Google’s Gemini Nano and Microsoft’s Phi Silica are similar. The future is local inference.

FAQ

Q: Is Apple Intelligence private? A: Mostly. On-device = private. Private Cloud Compute = claimed private but requires internet.

Q: Does it work offline? A: Yes for on-device features. No for Private Cloud Compute queries.

Q: Which Mac supports it? A: M1 and later. Needs 16GB RAM minimum.

Q: Is Foundation Model API free? A: Yes for app developers, no rate limits (yet).

WWDC26 Apple Intelligence: A Developer's Deep Dive After 3 Months