AI・テクノロジー
Apple Intelligence 2.0 Beta Released — On-Device Inference SLM Reshapes the Architecture of Privacy AI
機械翻訳 / Machine-translated
機械翻訳 / Machine-translated
On August 15, 2026, Apple activated the core features of "Apple Intelligence 2.0" in iOS 20 / macOS Tahoe Developer Beta 6. By integrating an inference-specialized SLM (small language model) that operates entirely on-device, the system handles step-by-step logical reasoning without ever sending data to the cloud. A new competitive axis — zero-transmission inference — has emerged to challenge enterprise AI, which has until now taken cloud dependency for granted.
In developer release notes dated August 15, Apple explicitly outlined the following:
Developers on X posted a stream of breaking observations:
"Apple Intelligence 2.0's Reasoning Mode is solving math benchmarks (GSM8K) on-device alone. This might open the door to processing medical records."
According to Apple's internal benchmarks, MMELU-based reasoning accuracy has improved by approximately 31% compared to the previous generation.
Apple officially unveiled the first-generation Apple Intelligence at WWDC25 in June 2025. Initially limited to lightweight tasks such as text summarization and notification organization, reasoning tasks depended on offloading to Private Cloud Compute. However, after Google's Gemini Nano 3 (March 2026) and Samsung's Galaxy AI 5.0 (May 2026) both shipped on-device inference as a standard feature, voices within the developer community questioning Apple's delayed response grew rapidly.
In regulated industries such as healthcare, legal, and finance, many enterprises maintain internal policies prohibiting data transmission to the cloud, making on-device AI less a feature differentiator and more a prerequisite for market entry. According to Gartner's Q2 2026 report, intent to adopt on-device AI among regulated-industry enterprises has expanded approximately 2.3 times year over year.
Apple has not disclosed parameter counts, but reverse engineering by multiple developers estimates the model at approximately 3 billion parameters. Custom Core ML quantization (4-bit) brings inference speed on the A18 Pro to approximately 45 tokens per second. Benchmarked against Meta's Llama 3.2 3B and Microsoft's Phi-4-mini, the model is considered likely to reach practical utility for summarizing and classifying legal and medical documents, depending on fine-tuning.
Article 13 of the EU AI Act mandates data traceability, and PCC 2.0's "immediate session deletion" design is consistent with that requirement. Friction with third-party transfer restrictions under Japan's amended Act on the Protection of Personal Information is also likely inapplicable under a cloud-free model, which could significantly shift legal teams' decisions about adoption.
Google has announced plans to launch Gemini Nano 4 within Q3 2026, and Samsung has already declared NPU enhancements for the Galaxy S26. The speed at which "on-device inference" is transitioning from a differentiator to a commodity is rapid, and the window in which Apple can maintain its lead from this update is estimated at roughly six to nine months.
What deserves attention is not the features themselves, but where they are being opened up. In this beta, a pathway has been confirmed through which third-party apps can call the inference SLM via the Core ML API. If that access is fully unlocked in the official September release, the groundwork will be in place for EMR (electronic medical record) applications and legal document review tools to complete AI processing entirely outside the cloud.
Reports have also reached us that multiple healthcare SaaS vendors have already begun private testing, suggesting that the emergence of AI applications tailored to regulated industries will accelerate heading into late 2026 and Q1 2027.
That said, risks must also be faced squarely. On-device models are bound to OS release cycles, meaning performance improvements on the scale of weeks — as is possible with cloud models — are not feasible. The upper limit of tasks a 3B-class model can handle remains unclear, and whether it meets acceptable error rates for multi-step financial calculations or complex legal documents is a question each organization will need to validate independently for the foreseeable future.
The question of "cloud AI or on-device AI?" has, over the past several months, shifted to "given that both are assumed, how do you combine them?" Apple Intelligence 2.0 has arrived as a concrete product that forces enterprises to make that architectural choice. The next focus is the scope of API access in the official September release — and which regulated industry will be the first to move to production adoption.
This article was written by an AI writer (AI News) from the Mirai News editorial team.
