Hugging Face AI Updates: October 30, 2026

1. Project ProtoStar Aims to Make Publicly Funded Science AI-Ready

Hugging Face announced Project ProtoStar, which pushes for publicly funded research datasets, models and code to be released in AI-ready, machine-actionable formats, and tests their value through open challenges. The founding university partners are Carnegie Mellon, Purdue, Georgia Tech, Arizona State and George Washington University, working through the Open Forum for AI, along with the Astera Institute and the ORCA research community. Hugging Face is the infrastructure partner. Hugging Face says it now hosts about 29,000 public scientific datasets, counting medicine and mathematics. New science datasets shared in the first nine months of 2026 were more than three times the number from the same period in 2025. The first challenge, CMU’s Climate Risk Intelligence Challenge, runs November 2 to 13. Teams use records through 2018 for 210 U.S. cities to forecast four climate indicators for 2019 to 2026, and final results are expected in early 2027. Source

2. AnySpeechRecognizer and AnyDecisionModel Bring Swappable Open Models to Apple Platforms

Hugging Face released two open-source Swift packages for macOS 26 and iOS 26 and later. AnySpeechRecognizer copies the shape of Apple’s SpeechAnalyzer API but lets developers swap Apple’s engine for NVIDIA Parakeet or Alibaba Qwen3-ASR running locally on Core ML. It also adds speaker diarization through Sortformer, which tracks up to four speakers. AnyDecisionModel returns typed answers with probabilities for yes/no, multiple-choice and rating questions, without generating free text, and runs on MLX, Core AI, llama.cpp, Ollama or a hosted Jev backend. On LibriSpeech, Parakeet Unified scored 1.43% / 2.19% word error rate (test-clean / test-other) against 2.12% / 3.97% for Apple’s transcriber on an M4 Max. It transcribes an hour of audio in roughly 20 to 40 seconds. Source

3. IBM Ships Granite Switch 4.1.1 With Updated Speech and Decision Adapters

Hugging Face. IBM published Granite Switch 4.1.1 on the Hugging Face blog. It is a backward-compatible update to Granite Switch 4.1, the modular architecture that serves specialized, cache-efficient adapters from one model endpoint. The 30B base model is unchanged, and only two components were retrained. The 470M granite-speech-5.0 ASR model now covers 10 languages and outputs native punctuation and capitalization. The granitelib-core Decision adapter improved on the Jev Decision Index test set, though IBM gives no scores. IBM pitches the release as updating models the way teams update software, and says an Adapter SDK for third-party adapters is coming. Source