WWDC 2026 (June 8–9) — impact on tinygpt + Pace

Researched 2026-06-10 via two web sweeps (platform + product). Sources at bottom; unverified items marked. macOS 27 is “Golden Gate”, fall 2026.

The three findings that change our roadmap

1. “Core AI” framework supersedes Core ML — and aims at our exact bottleneck

New inference framework (same runtime as Apple Intelligence): Swift API (AIModel, InferenceFunction, NDArray), Python toolchain (coreai-torch / coreai-opt / coreai-build), multi-function single-asset models, in-place mutable state views (no input/output state copies), AOT compilation, embedded custom Metal kernels, int4/int8/FP4/FP8 + palettized quantization, dedicated Instruments profiler.

Why it matters: M8’s decode ceiling is ~1.6 ms × 28 dispatches/token. A 28-function single Core AI asset + in-place state views is plausibly the first-party fix for both the dispatch overhead AND our IOSurface ping-pong hack. Also likely routes around the coremltools-9/macOS26 ANE-binding bug (no coremltools fix shipped — still 9.0 from Nov 2025).

Action (new M9 candidate): prototype 2-block Qwen3 chain as a 2-function Core AI asset; measure per-token dispatch vs M8; numerics gate applies. This replaces “retry M6 N≥2 on macOS 26” as the decode lever.

2. Foundation Models now runs CUSTOM models — Apple shipped our M8 idea

LanguageModel protocol: any model can back a LanguageModelSession. Apple ships open-source CoreAILanguageModel (ANE) and MLXLanguageModel (GPU) — the WWDC session literally demos custom Qwen3-0.6B/8B on ANE through it. Plus: rebuilt on-device model (8k ctx, image input, better tool calling), @Generable structured output, Spotlight RAG tool, Evaluations framework, fm CLI.

Why it matters: part-threat (structured gen + tool calling + free model commoditizes runtime surface), part-leverage (first-party ANE path for our exact base model). Our remaining moat: grammar-constrained decoding fidelity (FSM masks vs @Generable), the LoRA/DoRA adaptation loop, TTFW, and any-app AX control.

Action: benchmark CoreAILanguageModel(Qwen3-0.6B) vs M8 chain vs serve-int8 (tok/s, TTFW, constraint fidelity). Conform tinygpt’s serve to the LanguageModel protocol so Pace’s planner is model-pluggable either way.

3. Siri-Gemini + macOS 27 defines Pace’s launch window

Siri rebuilt on Gemini-derived models: screen awareness, multi-step cross-app actions via App Intents (SiriKit deprecated). BUT: waitlisted beta, M3+/12 GB requirement, English-only, fall 2026, App-Intents-bound (only apps that opted in), and Private Cloud Compute now runs on Google Cloud/NVIDIA infrastructure.

Why it matters: Apple just announced Pace’s category and simultaneously handed us the differentiation: any app (AX, no developer opt-in), any Apple Silicon Mac (M1+), zero cloud (“your voice never touches Google’s servers”), shipping now. The window is before fall 2026.

Secondary findings

• MLX first-party embrace: dedicated sessions, M5 neural-accelerator kernels (4× matmul vs M4), continuous batching in mlx-lm server, Thunderbolt-RDMA distributed inference (26.2). Validates our stack; pull M5 kernels + consider continuous batching for serve.
• Dictation in macOS 27: system-level auto-formatting, “hyper-accurate” local dictation (M3+/12 GB tier). Direct threat to the dictation specialist → re-scope Stage B toward custom vocab / app-aware formatting / voice-edit commands; treat dictation as feature, not wedge.
• Visual Intelligence on macOS (screenshot-based) + App Schemas: no general third-party computer-use API → the Pace Task Loop stays differentiated. Expose Pace skills as App Intents for Siri reach.
• Speech/TTS: nothing new at WWDC26; SpeechAnalyzer (WWDC25) still the comparison point. WhisperKit decision stands; re-bench on macOS 27 beta.
• Extensions framework: third-party AI providers (Gemini, Claude) selectable for Siri/Writing Tools. Monitor whether Pace can register as a provider. No new permission restrictions on AX/mic/screen found in beta-1 coverage (re-check at beta 2).

Roadmap deltas (concrete)

1. NEW research task: Core AI multi-function chain prototype (M9) — the decode-speed lever, supersedes the M6-retry idea.
2. NEW benchmark: CoreAILanguageModel vs M8 vs serve-int8 on Qwen3-0.6B.
3. Landing page: add the PCC-on-Google-Cloud contrast line + “any app, any Apple Silicon, today” vs Siri’s M3+/waitlist/fall.
4. Dictation Stage B re-scoped (custom vocab/voice-edit) — lower priority.
5. Launch clock: HN/public launch before fall 2026, ideally well before Golden Gate GA.
6. macOS 27 beta validation pass for Pace (AX intact, Intel dropped is fine).

Sources

Platform: Apple WWDC26 sessions 324/325/326 (Core AI), 241/339/298 (Foundation Models), 232/233/328 (MLX), 297 (Visual Intelligence), 240/345 (App Intents), 347 (agentic security); coremltools GitHub releases. Product: Apple Newsroom (next-gen Apple Intelligence), CNBC/TechCrunch/ Engadget WWDC26 coverage, MacRumors State of the Union, 9to5Mac (beta-1 waitlist), TechTimes (M3/12GB requirement), Neowin/digitimes (PCC on Google Cloud), macOS 27 developer release notes. Press-sourced items (Siri-Gemini details, “replaces Core ML” framing) are lower-confidence than Apple session pages.