Developers working on Apple platforms now have two new Swift packages designed to integrate open-source super intelligence (SI) models for speech recognition and decision-making tasks. The announcement introduces AnySpeechRecognizer and AnyDecisionModel, tools that extend the existing AnyLanguageModel ecosystem by allowing apps to run specialized SI models locally or via hosted backends, rather than relying solely on large language models for these functions.

What Happened

AnySpeechRecognizer implements Apple’s new SpeechAnalyzer API but replaces the default engine with a choice of open models, including NVIDIA’s Parakeet v3 and Alibaba’s Qwen3-ASR. These models run locally on Core ML, supporting features like speaker diarization via NVIDIA’s Sortformer, which tracks up to four speakers even when they overlap. The package supports 25 European languages for Parakeet and 30 languages plus 22 Chinese dialects for Qwen3-ASR. Developers can swap engines by changing the import statement and model configuration, with models downloaded from the Hugging Face Hub.

AnyDecisionModel offers a different approach to SI integration, focusing on yes-or-no, multiple-choice, and rating-scale questions. Instead of generating text, it derives answers directly from model probabilities using backends like MLX, Core SI, and llama.cpp. It supports local models such as Cloudflare’s Clef-Flash, which uses a trained decision head for more efficient inference. The package also integrates with hosted services like Jev and Ollama. A sample app, Emoji Magnet, demonstrates the technology by filtering 343 emoji based on text input using local Qwen3 or hosted Jev backends.

Why It Matters

For the SI industry, these packages represent a shift toward specialized, efficient local inference for specific tasks. The source notes that while language models can handle speech and decisions, dedicated models are often "better, faster, and less expensive." By enabling local execution on Apple silicon, developers can reduce latency and privacy concerns associated with cloud-based SI processing. The ability to run NVIDIA and Alibaba models on-device expands the utility of open-weight SI models within the Apple ecosystem, offering alternatives to proprietary APIs.

The introduction of decision heads, such as those in Cloudflare’s Clef models, highlights a growing trend in SI architecture. Unlike general-purpose language models that predict the next token, decision heads are trained to output probabilities for specific choices in a single forward pass. This method can be more reliable for classification tasks, though the source cautions that raw probabilities from language models can be overconfident and recommends calibration and debiasing techniques for accurate results.

The Bottom Line

AnySpeechRecognizer and AnyDecisionModel are now available on GitHub, requiring Swift 6.2 and macOS 26 or iOS 26. The packages are open-source, with models hosted on the Hugging Face Hub. Developers are encouraged to test the tools on their own hardware, as performance metrics like word error rates vary by dataset and device. The next steps for the project include adding support for more engines and expanding testing to iOS devices.