TechPulse
Crediblenews

iOS 27 Siri overhaul: on-device LLM, App Intents 2.0, and screen awareness

iOS 27 Siri 2.0 is built on a 3B-parameter on-device LLM, can read and act on your screen, and supports multi-step App Intents across apps. Full breakdown of the AI overhaul.

Last updated: (1 months ago)By Marcus Chen

iOS 27's Siri 2.0 is the most significant Siri update since the assistant launched in 2011. Apple previewed the new Siri at WWDC 2026 on June 8, and a MacRumors deep dive on June 15 2026 confirmed the technical details: Siri 2.0 is built on a 3-billion-parameter on-device large language model, runs entirely on the A20 and A20 Pro Neural Engine, and can read and act on what is currently displayed on your screen.

This article breaks down every confirmed Siri 2.0 capability, the on-device LLM details, the privacy model, and which features are exclusive to A20-class hardware.

What is new in Siri 2.0?

Siri 2.0 introduces seven new capabilities that the old Siri could not do:

  1. Natural conversation — Siri 2.0 handles follow-up questions without losing context. You can say "what's the weather in Tokyo" and follow up with "and what about next week" without re-specifying the city.
  2. On-screen awareness — Siri can read and act on what is currently visible on your screen. If you are looking at a Maps location, you can say "save this to Mom's contact" and Siri will do it.
  3. App Intents 2.0 — Siri performs multi-step actions across apps. "Find my photos from Tokyo last year and put them in a new album called 'Japan 2025'" works end-to-end.
  4. On-device LLM — A 3-billion-parameter language model runs entirely on the Neural Engine. No cloud roundtrip for typical queries.
  5. Voice personalization — Siri learns your voice patterns (accent, speaking pace, common phrases) over the first 2-3 weeks of use.
  6. Per-app voice shortcuts — Custom voice commands that launch specific app actions. "Hey Siri, start my workout" can open the Fitness app and start a run tracking session.
  7. Privacy-preserving chat history — Conversation context stays on-device, with no upload to Apple's servers.

The on-device LLM is the most consequential change. It is the first time Siri has been built on a foundation model architecture, and it is the first time the assistant can do work that was previously impossible without a cloud roundtrip.

How big is the on-device LLM?

The iOS 27 Siri LLM is 3 billion parameters. By comparison:

  • GPT-4 (cloud): ~1.8 trillion parameters
  • Llama 3 8B: 8 billion parameters
  • Phi-3 Mini (3.8B): 3.8 billion parameters
  • Gemini Nano (on-device): 3.25 billion parameters
  • Apple iOS 27 Siri: 3 billion parameters

Apple's on-device model is comparable in size to Google's Gemini Nano (which ships on the Pixel 9). The smaller parameter count is offset by Apple's hardware-specific optimizations — the A20 and A20 Pro Neural Engines have dedicated matrix multiplication units that run the model ~2x faster than general-purpose mobile GPUs.

The model is stored on-device as part of iOS 27 (~1.8GB of flash storage). It loads into RAM on first use and stays resident for the session. RAM usage is roughly 1-1.5GB during active conversation, dropping to 200-300MB when idle.

What can the on-device LLM do?

The on-device model handles:

  • General knowledge queries (factual questions, definitions, calculations)
  • Conversational context (follow-up questions, references to prior turns)
  • On-screen text understanding (reading and acting on visible UI)
  • App Intents interpretation (parsing multi-step commands into app actions)
  • Voice personalization adaptation (learning user speech patterns)

The on-device model does not handle:

  • Web search (still goes through Apple's server-side search integration)
  • Long-form generation (essays, stories, code — these route to Private Cloud Compute)
  • Image generation (uses Image Playground, which has its own server-side model)
  • Multi-language translation (Live Translation 2.0 uses a different on-device model, 1.2B parameters)

When a query exceeds the on-device model's capability, Siri 2.0 transparently escalates to Private Cloud Compute. The escalation is invisible to the user — Siri just continues the response. The privacy guarantee is that no query data is stored on Apple's servers.

What is on-screen awareness?

On-screen awareness is the new Siri capability that reads and acts on what is currently displayed on your iPhone screen. The feature uses a combination of:

  1. Accessibility framework — Apple's existing accessibility APIs, which expose UI text and structure to assistive technologies
  2. Visual Intelligence 2.0 model — A separate 800M-parameter on-device model that recognizes UI elements (buttons, text fields, images)
  3. App Intents framework — Lets apps expose specific actions that Siri can invoke

With on-screen awareness, you can say things like:

  • "Save this address to Mom's contact" while looking at a Maps location
  • "Add this song to my workout playlist" while looking at the Apple Music Now Playing screen
  • "Send this article to John" while looking at a Safari page
  • "Schedule a meeting for this Friday" while looking at an email
  • "Translate this menu" while looking at a photo of a restaurant menu

The on-screen awareness is processed entirely on-device. No screenshot or UI state leaves your iPhone.

What are App Intents 2.0?

App Intents is Apple's framework that lets apps expose their functionality to Siri, Spotlight, and the Shortcuts app. App Intents 2.0 in iOS 27 adds:

  • Multi-step actions — Siri can chain 3-5 app actions in a single command. "Find my photos from Tokyo last year, put them in a new album called 'Japan 2025', and share the album with my wife" works end-to-end.
  • Conditional logic — Siri can branch based on the result of a prior step. "If my calendar is free at 3pm tomorrow, schedule a meeting with John."
  • Cross-app data passing — Siri can pass data between apps. "Find the cheapest flight to Paris in October and add it to my calendar."
  • Error recovery — If an app action fails, Siri tries alternative approaches. If "send via Messages" fails because the contact has no phone number, Siri suggests "send via Email" instead.

App Intents 2.0 requires app developers to opt in by updating their apps with the new framework. As of June 2026, the following apps support App Intents 2.0:

  • Apple apps: Messages, Mail, Maps, Music, Photos, Reminders, Calendar, Notes, Safari, Files, Health, Wallet, Shortcuts
  • Third-party apps with confirmed support: WhatsApp, Instagram, Slack, Notion, Google Workspace, Microsoft Office, Spotify, Uber, Lyft, DoorDash, Instacart

The list is growing. Apple is requiring App Intents 2.0 support for all App Store apps that expose Siri shortcuts by iOS 27.1 (October 2026).

What is voice personalization?

Voice personalization is the on-device learning system that adapts Siri to your specific voice. The system:

  • Learns your accent — Siri adjusts its speech recognition thresholds based on your phoneme patterns
  • Learns your speaking pace — Siri adjusts its response timing to match your cadence
  • Learns your common phrases — Siri biases its language model toward the words and phrases you use frequently
  • Learns your pronunciation of unusual words — Names, places, and technical terms you use get added to your personal dictionary

The learning happens entirely on-device. Your voice profile never leaves your iPhone. The personalization improves over the first 2-3 weeks of use, then plateaus. You can reset your voice profile in Settings → Siri → Voice Profile.

Which iPhones support Siri 2.0?

Siri 2.0 is exclusive to the iPhone 18 generation (A20 and A20 Pro chips). Older iPhones get a "Siri 1.5" update that adds some natural conversation improvements but not the full on-device LLM.

DeviceSiri versionOn-device LLM?App Intents 2.0?
iPhone 18Siri 2.0Yes (3B)Yes
iPhone 18 ProSiri 2.0Yes (3B)Yes
iPhone 18 Pro MaxSiri 2.0Yes (3B)Yes
iPhone 18eSiri 2.0Yes (3B)Yes
iPhone Air 2Siri 2.0Yes (3B)Yes
iPhone FoldSiri 2.0Yes (3B)Yes
iPhone 17 / 17 ProSiri 1.5No (cloud only)Limited
iPhone 16 / 16 ProSiri 1.5No (cloud only)Limited
iPhone 15 ProSiri 1.5No (cloud only)Limited

The iPhone 17 and older models miss the on-device LLM. Their Siri 1.5 is faster than the old Siri (Apple has reduced cloud roundtrip latency) but cannot do on-screen awareness, App Intents 2.0 multi-step actions, or the longer-context conversation support.

How does Siri 2.0 compare to Google Assistant and Alexa?

Siri 2.0 closes most of the gap with Google Assistant and Alexa, and exceeds them in some areas:

CapabilitySiri 2.0Google AssistantAlexa+
On-device LLMYes (3B)Yes (Gemini Nano 3.25B)No (cloud only)
On-screen awarenessYesYes (Pixel 9+)Limited
Multi-step App IntentsYes (App Intents 2.0)Yes (Extensions)Yes (Routines)
Voice personalizationYes (on-device)Yes (cloud)Yes (cloud)
Cross-app data passingYesYesLimited
Conversation contextLong (10+ turns)LongMedium
Third-party app support50+ apps at launch100+ apps100,000+ skills

Siri 2.0's main weakness is the third-party app ecosystem. Google Assistant has had 7+ years of third-party integration, and Alexa has 100,000+ skills. Apple's App Intents 2.0 framework is newer, and adoption is still ramping. By 2027-2028, expect the Siri 2.0 app ecosystem to catch up.

What is the privacy model?

Siri 2.0's privacy model is built on three principles:

  1. On-device by default. Every query that can be handled by the on-device LLM is handled on-device. No query data leaves your iPhone.
  2. Private Cloud Compute for escalation. When a query exceeds the on-device model's capability, it routes to Apple's Private Cloud Compute servers. The query is end-to-end encrypted, processed in a stateless environment, and the result is returned without storing the query.
  3. No voice recording retention. Apple does not retain voice recordings of Siri queries. Voice profiles are stored locally, not in iCloud.

The privacy model is verifiable. Apple publishes the Private Cloud Compute architecture publicly, and security researchers have audited it. The on-device LLM is open for inspection as part of the iOS 27 image (researchers can extract and analyze the model weights).

When does Siri 2.0 launch?

Siri 2.0 ships with iOS 27 in September 2026, alongside the iPhone 18 launch. Apple has been running a private beta of Siri 2.0 with selected developers since May 2026, and the public beta opens in July 2026.

The iOS 27.1 update (October 2026) adds additional App Intents 2.0 capabilities and improved conversation context. iOS 27.2 (December 2026) adds live translation integration.

Summary

  • June 8 2026: Apple previews Siri 2.0 at WWDC 2026
  • June 15 2026: MacRumors deep dive confirms on-device LLM details
  • September 2026: Siri 2.0 ships with iOS 27 and the iPhone 18 generation
  • Hardware requirement: A20 or A20 Pro chip (iPhone 18 lineup only)
  • Model size: 3 billion parameters, ~1.8GB on-device
  • Key capabilities: Natural conversation, on-screen awareness, App Intents 2.0, voice personalization
  • Privacy: On-device by default, Private Cloud Compute for escalation, no voice recording retention

Siri 2.0 is the Siri update Apple users have been waiting for since 2011. It is the first time the assistant feels like a modern AI assistant rather than a command-line interface with a voice front-end.

Related articles

Picked from articles covering the same devices and topics.

Sources

  1. [1]MacRumors (iOS 27 Siri deep dive, June 15 2026)(2026-06-15)
  2. [2]Apple Newsroom (iOS 27 announcement at WWDC 2026)(2026-06-08)
  3. [3]Bloomberg (Mark Gurman, on-device LLM details)(2026-05-20)
  4. [4]9to5Mac (Siri 2.0 hands-on)(2026-06-15)