This article was originally published on runaihome.com
TL;DR: Apple's WWDC 2026 delivered the AI do-over it promised in 2024 — a 20B sparse on-device model (AFM 3 Core Advanced), a Gemini-powered Siri routed through Private Cloud Compute, and Xcode 27 agents. For home lab builders the practical change is narrow: if you write Swift apps, free on-device inference and provider-agnostic agents are real. If you run Ollama or llama.cpp, none of it touches your stack — a used RTX 3090 still wins on tokens/sec and model choice.
| Apple Foundation Models (AFM 3) | Open models on Apple Silicon | Used RTX 3090 + Ollama | |
|---|---|---|---|
| Best for | iOS/macOS app developers | 7B–70B open models on a Mac | Max tok/s, widest model choice |
| Cost | Free on-device inference | Device cost only | ~$1,070 used + ~$0.034/hr power |
| Speed | ~30 tok/s (20B sparse) | 28.4 tok/s on 70B Q4 (M4 Max) | ~95 tok/s on a 7B model |
| The catch | Apple's model only, 12GB+ RAM, Apple devices | 546 GB/s bandwidth ceiling | 24GB VRAM ceiling, 350W draw |
Honest take: WWDC 2026 makes Apple Silicon a better app development platform, not a better open-model platform. If you run Llama, Qwen, or Gemma locally, your RTX tower or Mac Studio works exactly as it did on June 7.
In our June 2 preview we laid out what the rumor mill expected from Apple's keynote. The keynote happened June 8, and most of it landed. This is the follow-up: what Apple actually shipped, with the real numbers, and a verdict for anyone who builds or runs AI at home rather than just shipping iOS apps.
What Apple actually announced
The headline is AFM 3 — the third generation of Apple Foundation Models. The model that matters for hardware discussion is AFM 3 Core Advanced: a 20-billion-parameter model that runs entirely on-device.
The interesting part is how it fits. AFM 3 Core Advanced uses a sparse architecture that activates only 1 to 4 billion parameters per request. Apple stores the full 20B weights in flash storage and loads just the relevant expert set into RAM once per prompt, through a lightweight dense routing block. That is a deliberate workaround to the RAM wall that has bottlenecked on-device models for years — you get the quality headroom of a 20B model without needing 20B parameters resident in DRAM.
Performance is roughly 30 tokens per second on iPhone 15 Pro and iPhone 17 Pro class hardware. That's the same ballpark as the previous 3B on-device model, which tells you the sparse routing is doing its job: a much larger model, similar latency.
The catch is the device floor. AFM 3 Core Advanced runs only on hardware with at least 12GB of RAM — iPhone Air, iPhone 17 Pro and Pro Max, iPads with M4 or later, Vision Pro (M5), and Macs with M3 or later. Older 8GB devices fall back to the smaller AFM 3 Core model. If your home server is an M1 Mac mini with 8GB, the flagship on-device model isn't for you.
The three-tier stack
Apple settled on a clear three-layer architecture, and it's worth understanding because it determines what touches your hardware and what doesn't:
- On-device (AFM 3 Core / Core Advanced) — expressive voices, dictation, on-screen awareness, structured extraction, and quick personal-context lookups. Runs on the Neural Engine and GPU of your Apple Silicon device. No network, no API key.
- Private Cloud Compute — heavier requests that still need Apple's privacy guarantees, run on Apple Silicon servers where Apple says data isn't stored or made readable.
- AFM Cloud Pro — the top tier for world-knowledge and complex reasoning. Apple says it matches Gemini Frontier quality and runs on NVIDIA GPUs in Google's cloud, custom-built in collaboration with Google's Gemini program.
So the new Siri is a hybrid: simple, personal, on-device work stays local; the chatbot-grade reasoning routes out to Gemini-class infrastructure. AppleInsider's reporting is worth noting here — Apple was explicit that the on-device models contain no Gemini weights. The Google collaboration lives in the cloud tier, not on your phone.
Foundation Models framework: the developer angle
For anyone writing apps, the framework got the upgrades that matter:
- Multimodal image input — you can now send images alongside text. The on-device model identifies objects, extracts text, and reads screenshots.
- A single API surface that unifies on-device, server-side, and third-party provider access. You can swap the underlying provider without rewriting your code.
- Open source this summer, with Linux server support — which is the genuinely surprising one, and the only WWDC item that reaches beyond Apple's own walled garden.
Xcode 27 agents
Xcode 27 ships a dual-engine agentic coding system: a local Neural Engine model for real-time Swift completion that never sends your source off-device, plus a cloud routing layer to Anthropic Claude, Google Gemini, or OpenAI GPT for heavier analysis. Xcode is now an MCP host (via a mcpbridge binary), so any agent that speaks the Model Context Protocol can read diagnostics, symbol info, SwiftUI previews, and the Swift REPL live. The agent can run test suites, drive the iOS Simulator through a new Device Hub, and pull crash reports from Organizer to fix the underlying code.
If your interest is AI coding rather than AI hardware, our sister site covers the agentic-IDE landscape in depth at aicoderscope.com — the Xcode 27 model is conceptually close to what Cursor and Cline already do, now first-party on the Mac.
What this changes for home lab builders (and what it doesn't)
Here's the part the keynote glosses over. There are two completely separate things people mean by "local AI on a Mac," and WWDC 2026 only moves one of them.
Track 1: Apple's own AI stack. Foundation Models, Core AI (Apple's modernized successor to Core ML), AFM 3, Siri. This is for shipping features inside iOS/macOS apps. It got materially better. The free on-device inference is real, the privacy story is solid, and the framework going open source could matter for cross-platform developers.
Track 2: running open-weight models yourself. Ollama, llama.cpp, LM Studio, vLLM, ComfyUI — Llama 4, Qwen3.6, Gemma 4, DeepSeek, Mistral. WWDC 2026 changed nothing here. Apple did not open the Neural Engine to third-party LLM runtimes, did not ship a faster Metal inference path as a headline feature, and AFM 3 is not a model you can pull into Ollama. Your ollama run workflow on June 17 is identical to June 7.
This distinction is the whole verdict. If you bought a Mac Studio to run Qwen3.6 and Llama 3.3 70B, the WWDC announcements are interesting news but not an upgrade to your rig.
The numbers that actually decide your hardware
For Track 2 — the thing this site is about — the bandwidth and tok/s reality is unchanged:
| Hardware | Memory bandwidth | Llama 3.3 70B Q4_K_M | 7B model | Notes |
|---|---|---|---|---|
| Mac Studio M4 Max | 546 GB/s | ~28.4 tok/s | ~87 tok/s | Unified memory, quiet, low power |
| Used RTX 3090 | 936 GB/s | offload needed (24GB) | ~95 tok/s | CUDA ecosystem, ~350W |
| AFM 3 Core Advanced | (flash-routed) | n/a — Apple model only | ~30 tok/s | 20B sparse, 12GB+ RAM |
A used RTX 3090 still has the highest memory bandwidth in this comparison at 936 GB/s, and bandwidth is what governs decode speed for local LLMs. In June 2026 it averages around $1,070 used (range $966–$1,189), which is remarkable staying power for a card this old — and a direct consequence of the GDDR7 shortage squeezing new GPU supply. The 24GB ceiling and ~350W draw (about $0.034/hour at $0.12/kWh) are the trade-offs.
The Mac Studio M4 Max at 546 GB/s wins on capacity and noise: its unified memory le
Top comments (0)