At WWDC26 in June, Apple introduced its third generation of foundation models — and the headline wasn't in the cloud. Alongside the expected server models, Apple announced two models that run entirely on the device: AFM 3 Core, the next generation of its 3-billion-parameter dense model, and AFM 3 Core Advanced, a 20-billion-parameter model with a sparse architecture that activates just 1 to 4 billion parameters per request (Apple Machine Learning Research, 9to5Mac).

Read that again: a 20B model, natively multimodal, living on a phone. For years the assumption was that anything serious had to round-trip to a data centre, and the device was just a thin client. The biggest consumer platform on earth has now committed to the opposite bet — the same bet we've been making with LyboAI Edge.

The clever part of AFM 3 Core Advanced is what it doesn't load. Instead of holding all 20 billion weights in active memory, the full model sits in flash storage (NAND), and the sparse architecture pulls in only the 1–4 billion parameters a given request actually needs. Think of it as a library with a small reading room: the whole collection is on the shelves, but only the relevant books come to the desk.

This matters well beyond Apple. It's a blueprint every device maker and every on-device runtime can learn from: capability scales with storage (which phones have lots of), not just with RAM and thermals (which they don't). Techniques like sparse activation, quantisation and distillation are how mid-range hardware — not just flagship phones — gets useful local intelligence over the next few years.

When we assess AI use cases with our VALUE-AI framework, three of the recurring wins for on-device execution are exactly the ones Apple's move underlines.

• Privacy by architecture, not by policy. When the model runs locally, personal context never has to leave the device to be useful. That's a stronger guarantee than any promise about what a server won't do with your data.
• Latency you can feel. No network round-trip means responses in the time it takes the silicon to think — and it works in a lift, on a plane, or in a paddock with no signal.
• Costs that don't scale with every message. Cloud inference bills grow with usage; local inference is effectively free once the model is on the device. For always-on companions, that changes which products are even viable.

None of this is new to anyone building at the edge — but it's different when the default assumption of the world's largest app ecosystem shifts underneath developers. On-device is becoming the baseline, not the exception.

The honest caveat in Apple's announcement: the heaviest lifting still escalates. Complex reasoning and agentic tool use are handled by AFM 3 Cloud Pro, Apple's most capable server model, running in its private compute environment (9to5Mac). Nobody — Apple included — is claiming every agent workload fits on a handset in 2026.

That's the pattern we expect to define the next few years, and it's how LyboAI Edge is designed: local first, escalate deliberately. The companion handles conversation, personal context and routine tasks on the device; when a job genuinely needs more horsepower or an external tool, the runtime hands off — explicitly, and on the user's terms. The architecture question for builders is no longer edge or cloud, it's where exactly is the line, and who controls it.

If you're building an AI companion or agent into your product, WWDC26 is a useful forcing function. Our practical advice, based on what we're building and shipping at LyboAI:

• Design local-first from day one. Decide which of your companion's jobs must work with zero connectivity, and treat the cloud as an escalation path rather than the default.
• Measure, don't assume. Small models punch above their weight on scoped tasks and fall over on open-ended ones. In Edge Studio we lean on evals for exactly this — run your real prompts against candidate models before you commit.
• Keep your stack portable. Apple's models serve Apple's platforms. If you want one companion across iOS, Android and desktop, build against a runtime layer you control rather than a single vendor's framework.
• Version and sign what ships. A model on a device is software you've distributed. Treat companions like releases — signed, versioned, and updatable — not like server configs you can silently change.

The direction of travel is set: the most personal AI will run where your life actually happens — on the device. That's the future we're building for with LyboAI Edge and Edge Studio. If you want to start experimenting with an on-device AI companion of your own, head to lyboai.app.