Apple used WWDC 2026 to introduce Core AI, a developer framework that replaces Core ML, and shipped five new foundation models alongside it. The headline: a 20-billion-parameter sparse model that runs entirely on your phone, activating only 1-4B parameters per prompt. That is a production first for dynamic sparsity in consumer hardware.
The Five-Model AFM 3 Lineup
Apple's third-generation foundation model family spans on-device and cloud tiers:
| Model | Where It Runs | Size | Active Params | Purpose |
|---|---|---|---|---|
| AFM 3 Core | On-device | 3B (dense) | 3B | Lightweight text, routing, fast NLU |
| AFM 3 Core Advanced | On-device | 20B (sparse) | 1-4B per prompt | New Siri, dictation, TTS, image understanding |
| AFM 3 Cloud | Private Cloud Compute | Undisclosed | — | Main cloud text and image understanding |
| ADM 3 Cloud | Private Cloud Compute | Undisclosed | — | Image generation (Playground, Reframe, Cleanup) |
| AFM 3 Cloud Pro | NVIDIA GPUs in Google Cloud | Undisclosed | — | Complex reasoning, agentic tool use |
The on-device models have disclosed parameter counts. The cloud models remain a black box on sizing.
Instruction-Following Pruning: How 20B Fits on a Phone
The technical breakthrough is Instruction-Following Pruning (IFP), originally published by Apple Research in January 2025. Instead of treating sparsity as a static architectural decision, IFP uses a small predictor that reads each prompt and dynamically selects which rows and columns of the feed-forward-network matrices to activate.
The results from the research paper were striking: a 3B activated model outperformed the 3B dense baseline by 5-8 absolute points on math and coding, matching the quality of a 9B dense model. Same active compute budget, roughly 3x better quality.
For the production deployment, Apple stores the full 20B model in flash (NAND), keeps shared experts in DRAM, and pages routed experts into DRAM only when the predictor selects them. This is how 20B fits in a phone without melting the battery.
Core AI Framework: Core ML's Successor
Core AI is the new developer framework for running LLMs and generative AI on Apple Silicon. Key capabilities:
- Replaces Core ML as the primary ML framework for neural language models
- Supports large language models and generative AI entirely on-device
- Foundation Models framework now accepts images (not just text)
- Includes
coreai-buildcommand-line tool for ahead-of-time model compilation - Core AI debug gauge and instruments for profiling inference performance
Core ML is not deprecated. Apps using decision trees, tabular feature engineering, or non-neural approaches should stay on Core ML.
What the Benchmarks Show (and Don't)
Apple's evaluation is side-by-side blind human preference against their 2025 baseline. There are no third-party benchmarks yet.
| Task | New Model Preferred | 2025 Baseline Preferred |
|---|---|---|
| Text (AFM 3 Core, on-device) | 45.6% | 23.3% |
| Text (AFM 3 Cloud) | 64.7% | 8.7% |
| Image understanding (AFM 3 Core) | >61% | — |
| Image understanding (AFM 3 Cloud) | 37.8% | 9.6% |
| Dictation quality (AFM 3 Core Advanced) | 44.7% | 17.6% |
AFM 3 Cloud Pro adds roughly 10% relative preference over Cloud on text, 14% on math, and 14% on image understanding.
What is missing: MMLU, SWE-bench, GPQA, or any standard benchmark. Apple does not compare against GPT-5.5, Claude Opus 4.8, Gemini 3.1 Pro, Qwen 3.7, or Llama 4. Every number is internal progress, not competitive positioning.
The Google Connection
AFM 3 Cloud Pro runs on NVIDIA GPUs hosted in Google Cloud and is refined using outputs from Google's Gemini frontier models. Craig Federighi was careful to distinguish "trained using" Gemini from "is" Gemini. The collaboration is real, but Apple owns the resulting model.
Availability Gaps
None of this works in the EU on iPhone or iPad at launch. Mainland China is also excluded. That limits the addressable developer audience significantly, especially for agentic applications where these markets are growing fastest.
What This Means for the AI Landscape
Apple's AFM 3 lineup is less about chasing frontier benchmarks and more about vertical integration. The 20B sparse on-device model is the real innovation: dynamic sparsity at consumer scale, shipping to millions of devices, with battery life that works. No other company has deployed this architecture at production scale.
For developers, Core AI opens new possibilities for privacy-preserving AI apps that run entirely on-device. The framework supports text and image input, making it viable for multimodal applications without cloud dependencies.
The open-weight community (DeepSeek V4.1, Qwen 3.7, Llama 4.5) still leads on raw benchmarks and transparency. Apple leads on deployment engineering. Different games, different scoreboards.
Key Takeaways
- Apple shipped 5 foundation models at WWDC 2026, led by a 20B sparse on-device model
- Core AI framework replaces Core ML for neural language model workloads
- Instruction-Following Pruning makes 20B viable on a phone by activating only 1-4B per prompt
- No third-party benchmarks yet, and no EU or China availability at launch
- AFM 3 Cloud Pro uses Google Cloud NVIDIA GPUs and Gemini-distilled training
- The real story is deployment hardening, not frontier model competition