NobodyWho enables on-device LLMs in Swift

NobodyWho has added a Swift-facing path for running language models locally on Apple devices, giving iPhone, Mac, visionOS and watchOS developers a package they can pull into apps without API keys or a network connection. The project’s Swift README says the library supports on-device AI inference for Swift apps and is powered by llama.cpp, with support for GGUF models from Hugging Face. (github.com) GitHub shows the Swift package can be installed through Swift Package Manager, and the project published a `nobodywho-swift-v1.0.0` release two days ago. (github.com) The release matters because it packages several features developers usually stitch together from separate components. (github.com) NobodyWho’s public starter examples for iOS and macOS list chat, tool calling, image and audio ingestion, and embeddings with retrieval-augmented generation and a cross-encoder reranker. The company’s GitHub organization also lists a separate `nobodywho-swift` repository described as on-device AI inference for Swift apps. ### What exactly did NobodyWho ship for Swift developers? The Swift README says developers can add `nobodywho-swift` to an Xcode project through Swift Package Manager and run models locally on iOS, macOS, visionOS and watchOS. (github.com) The same page says no API keys or network connection are required. GitHub’s release page shows the latest tagged Swift release as `nobodywho-swift-v1.0.0`, published two days ago, with a packaged `NobodyWhoNative.xcframework` asset. The main repository history also shows a recent “Feat/uniffi swift” change, indicating the Swift layer was being integrated into the broader codebase this week. (github.com) ### What can developers do with it besides basic chat? NobodyWho’s Swift starter examples say the package supports tool calling, including custom functions such as weather and calculator tools. The same examples list “Vision & Hearing” for image and audio ingestion with a multimodal model. (github.com) The starter apps also list embeddings and retrieval-augmented generation, including semantic search and a cross-encoder reranker. That means the Swift package is being presented not just as a chat wrapper but as a local inference stack that can retrieve documents, call functions and handle multimodal inputs inside an Apple app. (github.com) ### How does this fit Apple hardware? NobodyWho’s documentation does not make performance claims about Apple’s Neural Engine in the material surfaced here, but the package is explicitly aimed at Apple platforms and local execution. The Swift README says the software runs on Apple operating systems and uses llama.cpp under the hood, a widely used local inference runtime. (github.com) That setup gives developers a straightforward way to test private, offline model behavior on Apple hardware without routing prompts to a cloud API. The practical effect, based on the published feature set, is that a Swift team can prototype local assistants, document search, image-aware flows and function-calling agents inside native apps using a single package and example codebase. (github.com) That is an inference drawn from the documented features and platform support. ### What does this change for Apple-side engineering work? The published examples expand the surface area that must work reliably on device. Tool calling, audio ingestion, image ingestion and retrieval each add separate code paths, model choices and app permissions that teams have to validate across iPhone and Mac hardware. (github.com) That is an inference from the package’s documented capabilities rather than a statement by Apple or NobodyWho. The immediate next step is public and concrete: developers can inspect the Swift README, install the package from GitHub, and use the iOS and macOS starter examples released last week. (github.com) GitHub lists the latest Swift release as v1.0.0 and the starter-example release as v0.1.0, both available now.