Apple's Project Linwood & Greenwood

Apple is running a two‑track Siri overhaul to ship web‑answering and assistant upgrades quickly with third‑party LLMs (Glenwood) while converging on Apple Foundation Models under Private Cloud Compute for cost control, privacy guarantees, and long‑term independence (Linwood). ^{1 2 3}

The Glenwood initiative tests and deploys external frontier models (Google Gemini and Anthropic Claude) under private cloud control to deliver web‑answering and assistant features rapidly, whereas the Linwood program develops Apple Foundation Models (AFM) optimized for on‑device and Private Cloud Compute (PCC) inference to reduce unit costs and improve privacy.

The combination reflects a pragmatic strategy to ship generative AI features now while converging on in‑house models that meet quality and governance targets.

Apple's first attempt to re-architect Siri with generative AI under the codename Linwood began in 2023, but internal delays caused by technical shortcomings and quality gaps led to a public admission in March 2025 that some improvements would slip to 2026.

Bloomberg's reporting described a reorganization in which Mike Rockwell, formerly head of the Vision Pro program, took over Siri after CEO Tim Cook lost confidence in John Giannandrea's execution.

At the same time Apple faced competitive pressure as OpenAI, Google and other rivals were rolling out voice assistants powered by large-language models. To avoid a multiyear feature gap, Apple adopted a "two-track" strategy.

  Functionality stack

  Commercial and privacy mechanics

  Timeline

  Why this two‑track plan exists

Apple's own models are documented and deployable under PCC, but reporting indicates schedule slips and capability gaps versus frontier peers. The company therefore runs Glenwood in parallel to ship answers and assistant upgrades, while Linwood builds towards privacy‑preserving, lower‑cost, first‑party inference. Public sources describe both the reorganization to accelerate Siri and the active third‑party evaluations. ^{2 5 13}

  Operating model

Evidence for the planner/search/summarizer split and WKA, and for PCC's controls, appears in the linked reports and Apple's own security materials. ^{7 1 5}

  What to watch

  Assessment

The evidence supports a pragmatic plan: use Glenwood to reduce time‑to‑capability through a vendor model running on Apple servers, seeded by World Knowledge Answers, while pushing Linwood's AFM under PCC to reach parity, lower unit costs, and strengthen privacy posture. Bloomberg's reporting anchors the two‑track design, vendor tests, and WKA timing. Apple's own publications define AFM and PCC, including verifiable server images, research inspection, and non‑retention claims. This combination lets Apple ship now and own the stack later. ^{6 2 5}

  Primary sources

Bloomberg on Glenwood/Linwood, leadership changes, vendor tests, and WKA timing. The Verge on planner/search/summarizer and custom Gemini on Apple servers. Apple first‑party on AFM and PCC with security guarantees and inspection tooling. ^{6 7 2 5}

---

  Footnotes

1. Apple AI Search Plans, Bloomberg ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

2. Apple Foundation Models, Apple ML Research ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

3. Private Cloud Compute, Apple Security ↩ ↩²

4. Apple AI Turnaround Plan, Bloomberg ↩

5. PCC Documentation, Apple Security ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

6. Apple Considers Gemini AI, Bloomberg ↩ ↩² ↩³ ↩⁴

7. Apple AI Search Tool, The Verge ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

8. LLM Siri Early 2026, MacRumors ↩ ↩² ↩³

9. Apple Privacy Features ↩

10. Apple Intelligence Launch, Newsroom ↩

11. PCC Core Requirements, Apple Security ↩ ↩²

12. Rockwell Named Siri Head, Bloomberg ↩

13. Siri AI Overhaul Delays, Bloomberg ↩

14. AFM Tech Report 2025, Apple ML Research ↩