Apple hits 100k Duo apps as on‑device AI reshapes iOS

Apple announced that 100,000 Duo‑optimized iOS apps are now available, with thousands more arriving each day. The milestone marks the fastest adoption of on‑device AI on a mobile platform and forces developers to rethink how they ship intelligence.

Apple said in a developer briefing that the count hit 100,000 and that the pipeline adds new Duo‑compatible binaries daily. The figure comes directly from the company’s own metrics, not an estimate from third‑party analysts. Duo, Apple’s on‑device machine‑learning runtime, lets apps run models without sending data to the cloud, a promise that has become a selling point for privacy‑focused users.

The surge is more than a vanity metric. By moving inference to the handset, developers avoid latency spikes, reduce server costs, and sidestep regulatory scrutiny over data export. For engineers accustomed to cloud‑only pipelines, the shift means re‑architecting models for Apple’s Core ML format and testing on the limited memory and compute budgets of iPhone silicon.

On‑device AI gets personal: Notation’s chess coach

A small iOS app called Notation illustrates how on‑device AI can replace a cloud‑dependent workflow. The app bundles Stockfish 18, the strongest open‑source chess engine, and runs it entirely on the phone. It then hands the engine’s evaluation to a language model that translates raw scores into plain‑English coaching.

Notation’s price tag is £2.99 for launch, down from the original £6.99, and the purchase is a one‑time fee with no subscriptions or ads. The app’s description emphasizes that the LLM never evaluates positions itself; it only verbalizes a “complete coaching brief” prepared by the engine and a set of hand‑tuned classifiers. By keeping the entire pipeline local, Notation avoids the hallucinations that plague cloud‑based LLM coaches.

The product also tracks a user’s habits over months, building a “Shadow” persona that mimics past play styles. Because all the data stays on the device, the app can offer a private mirror for self‑analysis without ever uploading games. Notation’s approach shows that on‑device AI can deliver high‑quality, privacy‑preserving experiences that were previously thought to require server‑side horsepower.

Developers get an AI‑powered workflow: Async merges code, tasks, and reviews

On the developer side, the open‑source project Async stitches together Claude Code, Linear issue tracking, and GitHub pull‑request workflows into a single opinionated tool. After installing the Async GitHub App, the system imports open issues as tasks and spins up isolated Google Cloud Run jobs to analyze the repository.

The research phase runs a “research‑task” job that clones the repo, builds a code index, and generates clarifying questions. Once the developer approves, an “execute‑task” job launches Claude Code, which creates a feature branch, breaks the work into sub‑tasks, and commits each change separately. The resulting pull request appears in GitHub, and reviewers can step through focused diffs, request changes, or approve the batch.

Async’s architecture relies on isolated Cloud Run containers for every stage—research, execution, revision, and indexing—so the host codebase never sees stray state. The project is MIT‑licensed, and the repository includes scripts to spin up a matching Firebase config and pre‑commit lints. By automating the mundane parts of code review while keeping the heavy lifting on Google’s infrastructure, Async demonstrates how on‑device or tightly‑coupled AI can streamline mature codebases that are too risky for naive code‑generation tools.

What the Duo explosion means for the iOS ecosystem

The 100,000‑app milestone forces Apple to confront scaling challenges. Duo’s runtime must support a broader range of model sizes while keeping power draw low enough for everyday use. Early adopters have reported that some apps sacrifice model fidelity to stay within the 50 MB limit for on‑device bundles, a trade‑off that could push Apple to relax constraints in future OS releases.

At the same time, the market is rewarding privacy‑first solutions. Notation’s success shows that users will pay a premium for a coach that never leaves their device, while Async’s open‑source model proves that developers will adopt AI tools that respect the integrity of existing codebases. Both cases underscore a growing expectation: AI should enhance productivity without exposing data to the cloud.

Apple’s next move will likely involve tighter integration of Duo with Xcode, making model conversion and testing as seamless as adding a new framework. If the company can lower the barrier for developers to ship on‑device intelligence, the current growth curve could become a steep climb rather than a plateau.

What to watch next

The next iOS release is expected to include a Duo performance dashboard that reports model latency and memory usage per app. Keep an eye on Apple’s WWDC sessions for any changes to Core ML quotas or new hardware accelerators. For developers, the adoption rate of Async’s Claude Code integration will be a useful barometer of how quickly AI‑assisted coding moves from experimental to production. Finally, watch the pricing and feature updates of privacy‑centric apps like Notation; any shift toward subscription models could signal a broader monetization strategy for on‑device AI services.