Apple's Third-Generation Foundation Models (AFM 3): adaptive-compute MoE with per-prompt routing

TL;DR. Apple shipped its third generation of Apple Foundation Models (AFM 3) at WWDC 2026, a family of five models built in collaboration with Google, spanning on-device to Private Cloud Compute. The two on-device models are the headline for routing/efficiency: AFM 3 Core (a 3B dense model) and AFM 3 Core Advanced, a natively-multimodal 20B-total sparse MoE that activates only 1–4B parameters at a time depending on the request. Crucially, the routing decision is made once per prompt, early ("early routing"): a single up-front decision selects which experts will be active for the full model depth, and apparently also decides how many active parameters to allocate — adaptive compute at the prompt level. The server model (AFM Server) is trained and served with Parallel Track (PT) parallelism rather than expert parallelism (EP). Source detail comes from Apple's ML research blog, surfaced and unpacked by HuggingFace's @eliebakouch.

flowchart LR
  P[Prompt] --> ER{Early router<br/>one decision per prompt}
  ER -->|simple request| LOW[Activate ~1B params<br/>fewer experts]
  ER -->|complex request| HIGH[Activate up to 4B params<br/>more experts]
  LOW --> DEPTH[Same experts held<br/>for full model depth]
  HIGH --> DEPTH
  DEPTH --> OUT[On-device output<br/>20B total / 1-4B active]
  DENSE[AFM 3 Core: 3B dense] -.-> FAMILY[AFM 3 family of 5]
  SERVER[AFM Server: Parallel Track PT parallelism,<br/>sync once every few layers, not EP] -.-> FAMILY
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef decision fill:#fef3c7,stroke:#f59e0b,color:#78350f
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class P input
  class ER decision
  class LOW,HIGH,DEPTH,OUT output
  class DENSE,SERVER,FAMILY aux

Key points

Per-prompt adaptive compute. AFM 3 Core Advanced decides once, up front, which experts (and how many active params, 1–4B of 20B) to use for the entire forward pass. This is coarser than per-token MoE routing but cheaper to serve and predictable in latency, a deliberate on-device tradeoff.
Early routing fixes the experts for full depth, unlike standard per-layer per-token MoE. @eliebakouch notes it's unclear whether the early decision does layer selection or expert selection.
Parallel Track (PT) parallelism on the server. Think of PT as an extension of tensor parallelism (TP) that syncs once every few layers instead of multiple times per layer, cutting communication. Apple chose PT over expert parallelism (EP).
Five models, on-device to Private Cloud Compute, natively multimodal at the Advanced tier, built with Google.

How it relates to prior wiki knowledge

Product validation of the routing-as-policy thesis. The llm-routing page argues routing is becoming the central control surface. AFM 3 ships input-conditioned compute allocation into hundreds of millions of phones: the cheapest sufficient compute per request, decided up front. It is the consumer-scale counterpart to CLEAR's per-query compute rationing and Kappa-SwiGLU's confidence-adaptive MoE.
Lands the same day as Chiaroscuro Attention (06-09), which routes per token to cheap vs expensive operators. AFM 3 routes per prompt to cheap vs expensive expert sets. Research and product converging on the same input-conditioned-compute idea at different granularities.
Connects to the MoE muP scaling line (the Kurate-rated MoE scaling paper): a 20B MoE with 1–4B active is exactly the regime where stable-scaling parameterization matters.
The Private Cloud Compute / Google Cloud / NVIDIA-GPU angle ties to today's Industry Pulse (Apple expanding PCC to Google Cloud on NVIDIA GPUs).

Gaps

Apple's blog reports architecture, not independent benchmarks; quality vs. Gemma 4 / Qwen on-device peers is unverified.
Whether early routing's coarse per-prompt allocation costs accuracy on mixed-difficulty prompts (where per-token routing would adapt mid-sequence) is the open question.

Source: Twitter/X (@eliebakouch, Hugging Face) → Apple ML Research blog · raw: raw/twitter/2026-06-09-afternoon.json