Apple AFM3: The AI Breakthrough Bringing 20B Models to Phones

**David** (0:00)
Welcome to TechDailyai. I am your host, David, and joining me today is our expert guest, Sophia.

**Sophia** (0:05)
Hey, everyone, great to be here.

**David** (0:07)
You can sponsor this podcast for just $25.
Your message will be featured across major platforms like Apple Podcasts, Amazon Music, Spotify, and more. If you're interested, visit TechDailyai to get started today.

**Sophia** (0:20)
Such a great deal, honestly.

**David** (0:22)
Right. All right, so with that out of the way, I want you to imagine trying to fit like the entire New York Public Library into just an ordinary school backpack.

**Sophia** (0:32)
Oh, wow. Okay. That's quite the image.

**David** (0:34)
Yeah. Well, you want to carry all that knowledge around with you. Right. But the physical constraints just make it totally impossible.

**Sophia** (0:39)
Right. The bag would just tear open.

**David** (0:41)
Exactly. And that is essentially the fundamental physics problem that engineers have been wrestling with when trying to put really highly capable artificial intelligence natively onto a smartphone.

**Sophia** (0:52)
Yeah. The hardware just physically couldn't handle it.

**David** (0:54)
Until now, apparently. Because today, we are looking at how Apple claims to have finally cracked that exact problem with the unveiling of their third generation foundation models, or AFM3, which was announced in June 2026

**Sophia** (1:10)
It is honestly a remarkable architectural shift. I mean, we are looking at a complete restructuring of how these computational systems operate.

**David** (1:17)
Right. And doing it within a personal hardware ecosystem.

**Sophia** (1:20)
Exactly. And doing it without melting your device or draining your battery in 10 minutes.

**David** (1:25)
Which is the dream. And to you listening right now, if you are holding a modern smartphone, or if you rely on digital tools to manage your schedule, your writing or your photos, the stakes here are pretty monumental.

**Sophia** (1:38)
Oh, absolutely. They affect almost everything you do.

**David** (1:40)
Yeah. We are talking about these completely invisible frameworks that are going to fundamentally change everyday tasks. It really shifts the entire boundary between your local hardware, the cloud, and your personal privacy.

**Sophia** (1:52)
It's a massive paradigm shift for sure.

**David** (1:54)
Okay. Let's unpack this because starting with the broad architecture, we are basically looking at a family of five custom built foundation models.

**Sophia** (2:04)
Right. Yes. Five models. And they are strictly divided into two distinct domains.

**David** (2:09)
Two domains. Got it.

**Sophia** (2:10)
Precisely. So it's built as a hybrid ecosystem. On one side, you have the on-device models that are living right there on your local silicon.

**David** (2:17)
So that's the stuff actually running on the phone in your hand.

**Sophia** (2:20)
Exactly. And that includes AFM3 Core, which is a highly optimized, like 3 billion parameter baseline model. And then there's the AFM3 Core Advanced.

**David** (2:30)
Which is the big one, right?

**Sophia** (2:31)
Yeah. That one is an incredibly powerful 20 billion parameter multimodal model. But then on the other side of the equation, handling the tasks that the phone simply cannot physically process, you have the server-side models.

**David** (2:43)
Running in the cloud.

**Sophia** (2:44)
Right. Specifically running within Apple's private cloud compute infrastructure.

**David** (2:48)
And that cloud lineup is, well, it's very specialized. You have AFM3 Cloud serving as kind of the highly efficient workhorse.

**Sophia** (2:56)
Yeah. Handling the bulk of the complex text and reasoning.

**David** (2:59)
Right. And then there's ADM3 Cloud, which is dedicated natively to image generation and editing. And finally, you have AFM3 Cloud Pro for the absolute heaviest lifting, right?

**Sophia** (3:09)
Yes. Complex reasoning and agentic tool use. The really deep problem-solving stuff.

**David** (3:14)
Now, what really caught my eye here is the infrastructure behind that Cloud Pro model. Apple actually collaborated with Google and NVIDIA for this.

**Sophia** (3:22)
Yeah, which is huge news in the industry.

**David** (3:24)
Right. They are extending their private cloud compute to run on NVIDIA GPUs, like within Google Cloud. And I kept wondering, why would Apple, a company that is famous for designing their own world-class silicon, need to reach out to NVIDIA and Google?

**Sophia** (3:40)
Well, I think it highlights a very pragmatic acknowledgement of where the hardware industry actually stands today.

**David** (3:46)
How so?

**Sophia** (3:47)
Well, Apple's own silicon, the M-series chips powering their servers, they are phenomenal for edge computing, privacy-focused tasks, and standard server load.

**David** (3:55)
Well, they're super efficient.

**Sophia** (3:56)
Highly efficient. However, certain really complex, agentic AI tasks, like say, reasoning through a sprawling multi-step logic problem, they currently just benefit immensely from the raw, specialized throughput of NVIDIA's hardware infrastructure.

**David** (4:13)
Ah, so it's just raw horsepower.