0.4.2 — macOS/iOS 27 GA

Built and gated on the release macOS 27 (26A428) and Xcode 27 (27A266a) — the builds Apple shipped on September 14, 2026. No API changes: from: "0.4.1" resolvers pick this up automatically.

Start here: installation and smallest working example · Versioned ChatDemo

.package(url: "https://github.com/john-rocky/coreai-kit", exact: "0.4.2")

This is a patch release: the toolchain moved from the beta to the release builds, the requirement wording dropped "beta", and three catalog pins moved. The coreai-models runtime pin (0.2.4-zoo) is unchanged.

Fixed

• Gemma 4 E4B, 12B and 31B catalog pins move to re-saved decode bundles (gemma-4-E4B-CoreAI e9ba305a, Gemma-4-12B-CoreAI 266c0458, Gemma-4-31B-CoreAI 92bcdec5). The previously pinned bundles carried coreai-torch 0.4.0-era IR, which every OS 27 build since beta 2 refuses at AIModel.load (LLVM ERROR: cannot unwrap empty odiec_module_t) — re-measured on 26A428 on 2026-09-14. The re-saved bundles are the same weights with debug locations stripped and a coreai-core 1.0.0b2 producer stamp. All three answered through chat-cli from the new pins (E4B, 12B, and 31B — the 31B first run on an M4 Max takes about two hours of load and GPU specialization before the first token; budget for it).

Verified environment

• Mac: M4 Max, macOS 27.0 26A428, Xcode 27 27A266a, CI self-hosted on the same builds (.xcode-pin).
• Local gate on the release toolchain: swift build + swift test (86 tests, 15 suites) and the hybrid two-turn entry-check (Qwen3.5 0.8B; turn 2 recalls ORCHID).
• Independent clean clone of the 0.4.2 tag (f3c33b1): entry-check chat, fm (turn 2 recalls ORCHID), japanese, japanese-fm all PASS on the release toolchain, 2026-09-15 03:48 JST.
• Device rows in this train were measured on the iOS 27 RC 24A435. The iOS 27 release build is 24A437 and has not been re-measured yet.