Hugging Face AI Updates: June 24, 2026

1. IBM Research Releases CUGA Agent Examples on Hugging Face

Hugging Face. IBM Research published a post on the Hugging Face blog introducing cuga-apps, a collection of two dozen single-file working examples built on CUGA, an open-source agent harness that handles planning, the execution loop, tool calls, and state management. The examples are designed for developers to read, learn from, and adapt while focusing only on defining tools and prompts. Source

2. Hugging Face Automates Weekly huggingface_hub Releases

Hugging Face. The Hugging Face team detailed how it automated the huggingface_hub release workflow to ship weekly rather than every four to six weeks, using open-source tools and open-weights models with a human reviewer approving each publication. The system drafts release notes and announcements with AI, then applies deterministic verification before release, at a cost of roughly $0.25 per release through inference providers. Source

3. Transformers.js Pilots Cross-Origin Storage API for Shared Model Caching

Hugging Face. A post on the Hugging Face blog introduced an experimental Cross-Origin Storage API that lets web applications share cached files across different origins using cryptographic hashes instead of URLs. The approach allows AI models and WebAssembly runtimes to be downloaded once and reused across multiple websites, and Transformers.js has begun piloting it through an opt-in flag. Source

4. Local Open-Weight Models Triage GitHub Issues for OpenClaw Repo

Hugging Face. A Hugging Face blog post demonstrated using local open-weight models such as Gemma and Qwen to automatically classify GitHub issues and pull requests for the OpenClaw repository without cloud-based APIs. Running on consumer hardware via an agent framework with restricted shell access, the medium-sized models performed one-shot classification with reported good accuracy and no fine-tuning. Source