Hugging Face AI Updates: April 16, 2026

1. Kernels Promoted to First-Class Hub Repository Type

Hugging Face. Hugging Face officially launched Kernels as a browsable, first-class repository type on the Hub, joining Models, Datasets, and Spaces. The Kernels Hub provides 61 pre-compiled optimized GPU operators (attention, normalization, MoE routing, quantization) delivering 1.7-2.5x speedups over baseline PyTorch, supporting NVIDIA CUDA, AMD ROCm, Apple Metal, and Intel XPU. Source

2. VAKRA Benchmark Reveals Agent Reasoning and Tool Use Failure Modes

Hugging Face. IBM Research published an analysis of VAKRA, a tool-grounded executable benchmark for evaluating AI agents with 8,000+ locally hosted APIs across 62 domains and 3-7 step reasoning chains. Key findings show GPT-OSS-120B excels at API chaining, Gemini-3-flash-preview performs best on tool selection from large sets, and agent performance degrades significantly with reasoning hop depth. Source

3. HoloTab: AI Browser Companion Powered by Open-Source Holo3-35B

Hugging Face. HCompany announced HoloTab, a free Chrome extension powered by the Holo3-35B-A3B computer-use AI model that automates web-based tasks directly from the browser. Users can record routines, describe tasks in natural language, and the agent navigates websites autonomously without requiring technical skills. Source