Hugging Face has released two Swift packages, AnySpeechRecognizer and AnyDecisionModel, that let developers run open machine learning models on Apple devices. AnySpeechRecognizer mirrors Apple's SpeechAnalyzer API from macOS 26 and iOS 26, so an app can switch recognition engines by changing its import and transcriber setup. Supported engines include NVIDIA's Parakeet family, which runs on Core ML, and Alibaba's Qwen3-ASR, which covers 30 languages and 22 Chinese dialects and runs on the GPU. The package also adds speaker diarization to track who is speaking. AnyDecisionModel answers yes-or-no, multiple-choice and rating-scale questions. Rather than generating text, it derives typed answers and probabilities from the model's output, and it can run models through MLX, Core AI, llama.cpp or Ollama. Both packages download open models from the Hugging Face Hub only when an app requests them. Downloads range from about 250 MB for the Sortformer diarization model to between 0.9 and 2.4 GB for Qwen3-ASR, and each Parakeet model is about 600 MB. The converted Core ML models are published on the Hub with source checkpoints, per-file checksums and pinned revisions. For Australian businesses and professionals, this is developer tooling rather than a change to anything in everyday use. Teams building voice features into iOS or macOS apps gain a choice of speech engines, and on-device processing may simplify handling of voice data under the Privacy Act 1988, although legal obligations still depend on how each app collects, stores and shares that data. The main costs are storage for model downloads and development time, not per-call fees. Non-developers are unlikely to notice any difference, but teams already building voice features on Apple platforms may find it worth a trial.