Hugging Face AI Updates: September 2, 2026

1. Hugging Face Publishes 200+ WebGPU Kernels for Browser Inference

Hugging Face. The @huggingface/kernels JavaScript library loads 207 optimized WebGPU kernels from the Hub, covering matrix multiplication, normalization, convolution, and attention primitives. Developers install the npm package and call getKernel(), and the library handles shader compilation, memory allocation, and execution behind a standard interface. Hugging Face reports a 2.57x geometric mean speedup over ORT WebGPU across 809 test cases, and shipped Fleet, an in-browser benchmarking tool that crowdsources performance and correctness data across real devices. Source

2. BenchMIRT Finds Benchmarks Measure Something Other Than Their Stated Target

Allen Institute for AI. BenchMIRT applies multidimensional item response theory to individual benchmark questions, trained on results from 100 LLMs across 16 benchmarks and more than 34,000 questions. Two dominant dimensions emerged on their own, safety and general reasoning, and several benchmarks did not line up with their intended purpose: BBQ, built for social bias, correlated more with general reasoning, and WMDP scores tracked reasoning more than safety. Keeping only 10% to 50% of the questions preserved similar capability estimates, which suggests most benchmark suites are far larger than they need to be. Source