Apple Keeps 20 Billion-Parameter AI Dictation Off by Default in iOS 27 Beta

Q: Is Apple's 12GB RAM requirement for AI a technical need or a strategy to drive upgrades?

It is both a real technical limit and a product-segmentation choice, but the evidence points to a genuine hardware need at the core. Apple’s new dictation runs on-device using its 20-billion-parameter AFM 3 Core Advanced model. Although the model is sparse and activates only part of its parameters at a time, Apple still has to keep enough active memory available for routed experts, system processes, and low-latency inference. Apple says the model is stored in flash and dynamically pages needed components into DRAM using its Instruction-Following Pruning system, which strongly suggests memory pressure is a practical constraint, not just a marketing line. That said, the 12GB cutoff also clearly benefits Apple’s upgrade cycle. The feature is limited to newer premium devices, while the standard iPhone 17 with 8GB is excluded. That creates a visible gap between base and higher-end models, especially because dictation and Siri are everyday features users notice immediately. Even if the requirement is technically justified, Apple is also using it to differentiate its lineup. The strongest reading is that Apple chose a model ambitious enough to require 12GB for the experience it wants: fast, private, and fully local. It could likely have shipped a smaller or slower version to 8GB devices, but that would have meant lower quality or more cloud reliance. So the RAM requirement is not merely a sales tactic, yet it also conveniently pushes users toward newer, more expensive hardware.

3 articles · Updated · MacRumors · Jun 22

iOS 27 beta 1 leaves Apple’s new AI dictation disabled by default, even on supported devices including iPhone 17 Pro, iPhone Air and newer Macs and iPads.
Apple says the on-device system delivers a major accuracy gain: in human tests, AFM 3 Core Advanced beat the current dictation model 44.7% to 17.6% on overall quality and also led on punctuation, casing and meaning capture.
The 20-billion-parameter model fits on devices by storing the full model in flash memory rather than DRAM and activating only 1 billion to 4 billion parameters at a time through sparse expert routing.
That hardware demand limits the feature to devices with at least 12GB of RAM, excluding the standard iPhone 17 with 8GB; the same model also powers Apple’s opt-in expressive Siri voices preview.
Apple has not said whether the dictation upgrade will remain opt-in when iOS 27 launches later this year or be switched on during the beta cycle.