Apple's Advanced AI Dictation in iOS 27 Beta: What You Need to Know
Apple has been steadily pushing the boundaries of on-device artificial intelligence, and one of the most talked-about features arriving with iOS 27 is a dramatically upgraded AI-powered dictation system. However, if you've already jumped into the first developer beta and gone hunting for this feature, you may have noticed something unexpected: it isn't turned on by default. Here's a complete breakdown of what Apple's new advanced dictation is, how it works under the hood, which devices support it, and why it matters for the future of voice input on iPhone.
What Is Apple's New AI Dictation Feature?
Apple's next-generation dictation system is designed to deliver what the company describes as a "major boost in accuracy" over the existing dictation experience available on iPhones today. This isn't just a minor tune-up — it's a ground-up rethinking of how voice input is processed, corrected, and formatted in real time.
The key improvements center on more reliable on-the-fly capitalization and punctuation. Anyone who has used standard iOS dictation knows the frustration of watching a sentence come out with missing commas, incorrect capitalization, or awkward formatting. Apple's new system is specifically designed to address those pain points by using a far more powerful language model to interpret and format speech as it happens, rather than relying on simpler rule-based corrections after the fact.
The result is a dictation experience that feels significantly more natural and polished — one that can handle the nuances of everyday speech, including disfluencies, varied sentence structures, and stylistic differences between speakers.
Powered by the AFM 3 Core Advanced Model
At the heart of this new dictation system is Apple's AFM 3 Core Advanced model — the third generation of Apple's own foundation models and one of the most ambitious on-device AI systems the company has ever built. Understanding a little about this model helps explain both why the feature is so impressive and why it comes with a few hardware limitations.
AFM 3 Core Advanced is a 20-billion-parameter model, making it significantly larger and more capable than the models powering previous Apple Intelligence features. It is natively multimodal, meaning it's built from the ground up to handle different types of inputs — not just text, but also signals and context from other modalities — rather than having multimodal capabilities bolted on after the fact.
Despite its scale, the model uses a sparse architecture, which is a critical design choice that makes it feasible to run on a smartphone at all. Rather than activating all 20 billion parameters for every single task, the model selectively activates just one to four billion parameters at a time depending on the nature of the request. This dramatically reduces the computational load while preserving the model's overall intelligence and capability.
How Apple Fits a 20-Billion-Parameter Model on a Smartphone
Getting a model of this size to run natively on a mobile device is an engineering challenge that Apple has tackled in a distinctly creative way. The full AFM 3 Core Advanced model is stored in flash memory rather than in DRAM, the faster but more limited memory traditionally used for active processing. This allows the model to physically reside on the device without requiring an impractical amount of high-speed RAM.
To make this work efficiently during real-time tasks like dictation, Apple uses a lightweight routing block that selects a fixed set of "experts" — specialized subnetworks within the larger model — during initial processing. These experts are periodically reselected during generation as the context of the task evolves. Apple calls this technique Instruction-Following Pruning, and it's central to making the model both fast enough and accurate enough for a live dictation scenario where latency is critical.
This approach represents a meaningful step forward in how large language models can be deployed at the edge — on devices rather than in the cloud — with real implications for user privacy, since audio and processing remain entirely on-device.
Human Evaluations Show Clear Performance Gains
Apple hasn't just made internal claims about the improvement. The company conducted side-by-side human evaluations comparing AFM 3 Core Advanced against its previous production dictation system across seven quality dimensions: overall quality, punctuation, casing, layout, meaning capture, disfluency handling, and style.
The results were striking. On overall quality, human evaluators preferred the new AFM 3 Core Advanced output by a margin of 44.7% to 17.6% over the prior system. Crucially, that preference held consistently across all six remaining quality dimensions — not just one or two. This kind of broad, consistent improvement across multiple evaluation criteria suggests that the gains are genuine and systemic rather than cherry-picked for a narrow use case.
Which Devices Support the New Dictation?
Because of the model's substantial size, Apple's upgraded dictation system is limited to a small number of newer devices. Currently, this includes the iPhone 17 Pro and the iPhone Air — devices with the processing power and memory architecture required to run the AFM 3 Core Advanced model on-device. Older iPhones, even those that support other Apple Intelligence features, are not compatible with this particular capability.
This hardware restriction is likely to be a point of discussion among Apple users, as it draws a clear line between which devices get the full next-generation AI experience and which do not.
Why Isn't It Enabled by Default in the Beta?
The fact that this feature is not toggled on out of the box in the first iOS 27 developer beta is consistent with how Apple typically handles experimental or computationally intensive capabilities during early testing phases. Developer betas are primarily intended to help app developers test compatibility and explore APIs rather than to deliver a polished end-user experience. Keeping a resource-heavy feature like advanced AI dictation off by default in beta reduces the risk of impacting device performance or battery life in unintended ways while Apple continues to refine the system.
As iOS 27 matures through subsequent beta releases over the summer, it's reasonable to expect Apple to enable the feature more prominently — potentially making it opt-in for users during setup, or even the default behavior on supported hardware by the time the public release arrives in the fall.
The Bigger Picture: On-Device AI Is Getting Serious
Apple's advanced AI dictation in iOS 27 is more than just a better voice-to-text tool. It's a signal of how seriously Apple is investing in large, capable AI models that run entirely on the device. By combining sparse model architectures, flash memory storage, and intelligent routing techniques like Instruction-Following Pruning, Apple is charting a path toward genuinely powerful AI experiences that don't require a constant cloud connection — keeping user data private while still delivering results that rival server-side systems.
For iPhone 17 Pro and iPhone Air users, the arrival of the full iOS 27 public release this fall promises a dictation experience that is meaningfully smarter, more accurate, and better suited to the way people actually speak. Whether you're composing emails, drafting messages, or taking quick notes hands-free, the leap in quality from the current system to AFM 3 Core Advanced is one of the most tangible AI upgrades Apple has shipped in years.