Hugging Face releases open Swift packages for on-device speech and decisions on Apple platforms

What was released

The Hugging Face blog announced on 29 October 2026 that mattt had released AnySpeechRecognizer and AnyDecisionModel, two open-source Swift packages extending the AnyLanguageModel approach to speech and decisions. The packages mirror the team's earlier 1.x releases of AnyLanguageModel and swift-huggingface: a familiar Swift API with your choice of model, all running locally.The design goal is explicit: for a customer support call, you first want speech recognition (what did they say) and then a decision (where do they go). Rather than handing both to a language model, the packages use models built for each task that are better, faster and cheaper.

AnySpeechRecognizer

AnySpeechRecognizer implements Apple's new speech-to-text API — SpeechAnalyzer, introduced in macOS 26 and iOS 26 — with a choice of recognition engine. Apple's Speech framework stays the default, but developers can drop in NVIDIA's Parakeet or Alibaba's Qwen3-ASR running on the package's own Core ML runtime.

The package also adds speaker diarization to track who speaks when, plus word timing and word confidence for live and file input. Qwen3-ASR is available in 0.6B and 1.7B sizes, recognizes 30 languages and 22 Chinese dialects including Mandarin, Cantonese, Japanese and Korean, and auto-detects the language — coverage Parakeet does not provide. Importing AnySpeechRecognizer instead of Speech is enough for most apps since the package mirrors the analyzer, transcriber, detector and asset inventory API.

AnyDecisionModel

AnyDecisionModel answers yes-or-no, multiple-choice and rating-scale questions with typed results and probabilities. Instead of generating text, it derives each answer from the model's probabilities — a more reliable and cheaper path for classification-style decisions.

It runs a model on device with MLX, Core AI or llama.cpp, or can point at a decision model from Jev or Ollama, and downloads open models from the Hugging Face Hub. Because the answer comes from logits rather than text generation, the result is deterministic relative to the model and includes a confidence score.

What it means

These releases continue a pattern visible across 2026: on-device AI is becoming a model market rather than a one-vendor feature. Apple's SpeechAnalyzer ships with Apple's engine; AnySpeechRecognizer keeps the API and swaps the engine. Developers get privacy, no per-call cost and offline operation while keeping the user experience identical.

For decision use cases, reading probabilities instead of generating text is a meaningful engineering improvement: no hallucinated preamble, no parsing, just a typed answer plus a confidence value. Combined with the Hub download path, this makes it practical to ship small, task-specific open models inside macOS and iOS apps.

Caveats and what to watch

AnyDecisionModel is new and the Jev/Ollama integration and supported decision-model formats will mature over time; the blog does not publish benchmark numbers for either package. On-device efficiency depends on the Apple silicon and chosen model size.

Watch adoption of the packages in shipping apps, the quality of Qwen3-ASR's 30-language coverage in real use, and whether Apple's own SpeechAnalyzer gains engine choice or diarization natively. Watch whether the logits-based approach becomes the default for classification-style model calls on Apple platforms.

The AnyLanguageModel lineage

The two new packages are part of the wider AnyLanguageModel and swift-huggingface 1.x ecosystem, which the blog summarizes in its AnyLanguageModel 1.0 recap: a year of local and hosted models in Swift. The unifying idea is that Swift developers should get one familiar API and a choice of model — local first, hosted when needed.

That positioning matters for the broader trend: the Apple developer platform is where consumer AI features actually reach users, and open-source Swift wrappers like these compete directly with vendor-locked SDKs by making model choice a drop-in decision. For independent developers, that lowers the cost of shipping on-device AI features without per-call cloud bills.

Why on-device matters now

On-device inference has moved from a niche to a default expectation, driven by privacy requirements, latency and cost. Speech recognition for dictation, transcription and accessibility runs continuously in many apps; sending every utterance to the cloud is expensive and slow, and for sensitive audio it is a privacy liability.

The packages lean into that shift: models download from the Hub once and then run locally, with no network dependency at inference time. The Core ML and MLX runtimes mean the same Swift API works across Apple silicon on both Macs and iPhones, and llama.cpp covers additional model formats. The result is a genuinely portable local-AI stack for Apple apps.

Caveats

The blog publishes no benchmark numbers, so the real-world quality of Qwen3-ASR speech accuracy and the decision models' calibration is unverified here — developer experience will be the test. The package list, model choices and available engines can change quickly as the ecosystem matures.

'Decision model from Jev or Ollama' integrations are new and their stability is unproven. AnyDecisionModel's logits-based answers are deterministic for a given model but cannot solve reasoning tasks the way a general language model can — the tradeoff is speed and reliability for narrow classification-style decisions.