MaiaEmbedded 1.1.0

MaiaEmbedded 1.1.0 is a comprehensive public-documentation update. Inference behavior and existing public API signatures are unchanged.

• Rewrite the README, testing and example guides for package consumers; remove obsolete implementation, release sign-off and personal-workflow material.
• Replace the experiment write-up with a physical-hardware benchmark guide. Document the Mac as Apple M5 Max, 18 CPU cores and 48 GiB RAM; preserve the original numeric measurements and remove simulator results and development-state reports.
• Expand Quick Help comments across the direct API, UCI adapter and benchmark support, covering inputs, ratings/history, sampling/RNGs, cancellation, callback queues, readiness and resource cleanup.
• Add a benchmark-support DocC catalog, expand the existing catalogs and provide complete Swift examples with deadlines and cleanup on success/failure.

Validation: three DocC catalogs compiled without warnings; all five complete Swift guide examples compiled in a separate client, with real-model direct/two-rating/sampling/UCI checks and failure/cancellation cleanup. The full local gate passed, including 23 tests each in Debug, Release and Thread Sanitizer, CLI/smoke/soak, benchmark and platform builds. Markdown links, benchmark summary rounding and unchanged original measurements were verified. Non-comment Swift code is identical to 1.0.0.

Source-only release; no prebuilt app, framework, model or runtime binary is attached. Swift Package Index discovers the tag automatically.

MaiaEmbedded 1.0.0

MaiaEmbedded 1.0.0 is the first public source release of native Maia-3 79M inference for iOS, iPadOS and macOS.

• Direct Swift API for human-move probabilities, rating-conditioned predictions, temperature/TopP selection, history and model lifecycle.
• Restartable in-process UCI adapter over the same core, plus desktop UCI, smoke and soak tools.
• Versioned public SwiftChessTools dependency, pinned ONNX Runtime, explicit SPI build configuration and DocC for both library products.
• Separately provisioned, size/SHA-256 verified 79M FP32 model; no model download during inference and no weights committed.

Requires Swift 6.2+, iOS/iPadOS 26+ or macOS 26+. Local validation used Xcode 27.1 / Swift 6.4 and Apple silicon. Intel is unqualified. Package/model licensing is AGPL-3.0-only; retain upstream and dependency notices.

Validation: full Debug/Release/Thread Sanitizer package tests (23 per configuration), real-model corpus and lifecycle, UCI CLI, smoke/soak, benchmark, DocC, package-only SPI schemes and platform app builds. Standalone asset-free source and separate direct/UCI consumers passed. Existing physical Mac/iPhone/iPad correctness/performance and user gameplay/lifecycle evidence is documented in Docs/Testing.md; native inference source/runtime/model are unchanged by release packaging.

Source-only release. No prebuilt app, framework, model or runtime artifact is attached. Provision the model with Scripts/download-model.sh; see README and Docs/Provenance.md.