Siri AI

It’s been challenging for me to think and write about Siri AI this summer, and I’ve already explained why: as a person whose workflow now revolves around agentic tools, I was not impressed by Siri AI at first.

I don’t think I was the only “power user” on Apple platforms who’s had this thought since getting their hands on the first beta of iOS 27 in June. If you’re reading this, chances are you’re pretty well versed in what models by OpenAI, Anthropic, or others now offer in terms of assistant features, web-based app integrations, memory, long-running tasks, and more. The product surface of Siri AI – the tentpole feature of iOS 27 that Apple is finally shipping after a two-year delay – has very little in common with those other agents or assistants, whatever you want to call them. Technically, yes – they are all based on a large language model under the hood. From a benchmark perspective, Apple’s new Foundation models are vastly behind the latest state of the AI frontier currently occupied by four American labs.

But as a product for people, Siri AI has a threefold advantage that no other company can match:

1. It’s built right into your phone or computer
2. It has a built-in knowledge graph about you in the form of personal context
3. It can act upon, and retrieve data from, the native apps you use every day, from Apple to third-party ones

And these advantages sum up why I struggled to understand Siri AI at first. Is Siri AI only passable because it has an unfair advantage due to proprietary integrations that other AI models can’t use? Can it stand on its own as a good implementation of AI for most people? Does it fundamentally reinvent how we use our devices, or is it just a lot of hype for a slightly better Siri?

Unpacking the answers to these questions will be my goal with this chapter. But I’m going to do something unusual and share my takeaways up front, since they frame how I’ve been using and thinking about Siri AI.

I find Siri AI’s underlying models to be subpar compared to the competition: for anything beyond a simple knowledge query, I can’t see Siri AI replacing ChatGPT and its superior reasoning modes combined with web search. For simpler queries – the kind that a Google AI overview could also capably answer – the balance of speed and terse formatting is optimal in Siri AI. If I’m watching a movie and want to know how old an actor is, or if I want to know the meaning of a word, Siri AI is perfect for it and I no longer need to reach for ChatGPT or my personal agent in Open Minis. As we’ll see later, I find the Siri app to be extremely barebones and representative of the state of the art for chatbot design…two years ago.

The lack of an “agent mode” for Siri AI stings, and it’s why I can’t replace my beloved Codex or Hermes Agent setups with it.

However, Siri AI eventually clicked for me for the opposite reason: it can do things in and with apps that no other third-party agent can. Yes, they are simpler actions, and, no, results aren’t always perfect; but if you love the Apple app ecosystem and prefer using excellent native apps to cross-platform web ones, Siri AI is a breath of fresh air in the AI landscape right now.

See, Siri AI is complicated, polarizing, and almost impossible to agree upon.

Welcome to AI, Apple.

Foundation Models and Siri AI’s Architecture

Before I cover what the new Siri actually looks like and how it is split across a system-wide assistant and a brand new app, let’s talk about what it is and what powers it under the hood.

The new Siri is powered by a system orchestrator that coordinates user requests across different modalities (text, voice, and images) and five new Apple Foundation models, which are:

• AFM 3 Core
• AFM 3 Core Advanced
• ADM 3 Cloud
• AFM 3 Cloud
• AFM 3 Cloud Pro

The first two are on-device models, with Core Advanced being exclusive to more recent devices with higher RAM. ADM Cloud is an image diffusion model that runs in the cloud, and which I will not cover in this review since I have no interest in generative images. AFM Cloud and Cloud Pro are cloud models running on Private Cloud Compute, with some differences: while Cloud is optimized for latency and efficiency (both for users, since it’s faster, and for Apple, since it costs less to serve), Cloud Pro is capable of deeper reasoning and agentic tool use, and it does not currently run on Apple-made servers. Instead, Apple teamed up with Google and NVIDIA to host Cloud Pro on NVIDIA GPUs running on Google Cloud, while maintaining the same privacy and security guarantees as “regular” PCC. (Only Apple can deploy software to these nodes, and only Apple can sign that software. More details here.)

All of these models were, according to Apple, trained with proprietary data and reinforcement learning, then “refined” with data from frontier Gemini models. Essentially, the new AFM family of models was refined with distilled Gemini outputs. I can’t stress this enough: although Siri AI and AFM may exhibit some behaviors reminiscent of Gemini (as I’ll explain later), Siri AI is not “powered by Gemini” and there is no Gemini to be found in Apple’s operating systems. No Gemini model or agentic harness is being used by Apple.

There are a couple more details worth exploring about the underlying nature of Siri AI. The first concerns what is, in my opinion, the most interesting model of the bunch – the new AFM 3 Core Advanced that runs on-device. For starters, Apple now has a larger local model that is multimodal: Core Advanced supports vision capabilities and powers the advanced dictation and expressive voices of the new Siri. Because of the increased footprint and technical requirements of this model, however, Apple had to restrict it to more modern devices:

• iPhone Air, iPhone 17 Pro, iPhone 17 Pro Max, iPad (M4) or later with at least 12 GB of RAM
• Mac (M3) or later with at least 12 GB of RAM
• Apple Vision Pro with M5 (I always forget this exists)

The architecture of the model itself is also interesting. As I previously explained, while AFM 3 Core is still a 3B model with a dense architecture, AFM 3 Core Advanced is a 20B model that can run on devices with 12 GB of RAM. That’s made possible by a technology that Apple invented for sparse architecture models in which the model is stored in flash memory, activated parameters are locked in per prompt, and only 1 to 4 billion parameters are activated concurrently depending on the user’s request. From Apple:

One area of deep innovation is our most powerful on-device model, AFM 3 Core Advanced. Traditional large language models—whether dense or sparsely activated—require all weights to reside in active memory (DRAM), creating a massive footprint that limits scalability on consumer hardware. To break this barrier, AFM 3 Core Advanced introduces a novel sparsely activated architecture built on Instruction-Following Pruning (IFP), a technique developed by Apple researchers.
Instead of forcing the entire model into DRAM, the full model is stored in flash memory (NAND). Because NAND-to-DRAM bandwidth is too slow to swap weights token by token, as standard MoE models require, AFM 3 Core Advanced makes routing decisions per prompt. A lightweight, dense block selects a fixed set of experts during initial processing, periodically reselecting them during generation. To minimize data movement, the model relies on a high percentage of always-active “shared experts” alongside input-dependent “routed experts” swapped into DRAM only when needed.

You don’t need to be an AI expert to understand that what Apple wants to achieve here seems remarkably complex and optimistic. Here we have a company that is two years late to its original promises for an AI-powered Siri, and which is saying it has built a system to orchestrate five models, across multiple modalities, with support for proprietary World Knowledge (since you can now ask Siri general questions about anything, too), and which is going to draw from your personal context on-device as well as integrate with all kinds of third-party apps. And while all of this is happening, the system should know how to split inference across local and cloud models, how to allocate reasoning effort levels, how to coordinate actions across apps, and how to do it all in less than 15 seconds with a nice, minimalistic UI.

When I started working on this review, I struggled to visualize and understand the new Siri AI architecture in my head, so I built something with Fable 5.1 and GPT-6 Astra to help me, and I’m embedding it in this review because maybe it’ll help you too. For the following visualization, I started from Apple’s official Siri AI graphic that the company shared at WWDC, then I combined it with the Apple Intelligence Report that can be exported directly from the Settings app in iOS and iPadOS 27. The report contains all Siri AI requests that were issued by the user over the span of 7 or 30 days, and it includes details on the steps taken by Siri AI, what tools were loaded, how many tokens were in the system prompt, what third-party apps returned via App Intents, how much time each step took, and more.

The interactive visualization shows you some examples from my real requests – typos and everything – that were performed with Apple’s built-in apps as well as AskPlay and Emoji Countdown. As you can see, it’s a very intricate system that is executed remarkably fast by Siri AI and the system orchestrator.

By and large, what Apple has built works, and as a user of Apple devices and the new Siri AI, you will never see any of its underlying complexity.

The new Siri AI is far from perfect and objectively behind the state of frontier AI models, but it’s also doing things that no other company is, and I firmly believe it is going to be good enough for millions of people.

If you’re coming from any other AI product from OpenAI, Anthropic, Google, or the rest of the industry, you must reset your expectations in more ways than one when it comes to Siri AI.

First, unless you’ve been on the beta and have already let your device go through its new indexing process, you’ll probably have to wait a few days until you a) get access to Siri AI and b) can use it to its full extent. In order to work properly, Siri AI needs to build an index of your context and apps, and this process isn’t quick. You can speed up the indexing by keeping your phone on Wi-Fi and connected to a charger, but it’s not immediate. You can monitor its progress via a system message displayed in the Settings app; in my experience, while the initial indexing takes a while, subsequent updates to the index are much faster, if not downright immediate in some cases (such as downloading a new app and asking Siri AI questions about it).

Second, you should let go of a bunch of UI conventions you’ve probably grown accustomed to when using ChatGPT or Claude over the past two years. There are no model pickers, reasoning selectors, skills, @mentions of plugins, context windows, memories, “fast modes”, or usage resets in Siri AI. This new product embodies the quintessential “it just works” Apple mantra – sometimes to a fault. The new Siri looks and works like a more conversational Siri that also comes with a dedicated Siri chatbot app. There is nothing you can configure or tweak about it besides, well, using your devices with different apps.

There is a fascinating dichotomy occurring between the underlying complexity of Siri AI and the utter simplicity presented to users. Most frontier AI models and products wear their complexity on their sleeves, with different levers to pull and settings to tweak if you want to get the most out of them. Apple isn’t giving users this kind of customization (or responsibility?) yet, and I think it’s the correct choice to ease over two billion people into the uncomfortable world of AI.