Ant Group and vLLM Community Open-Source Humming Kernel Library

Ant Group and the vLLM community have released Humming, an open-source kernel library designed for quantized AI inference across multiple GPU types and CUDA-compatible devices.

Humming’s unified architecture supports combinations of weight and activation formats for dense and mixture-of-experts (MoE) models. It covers multiple integer and floating-point types, bit widths and scaling methods.

The library’s execution performance significantly improves on vLLM Marlin, FlashInfer and DeepGEMM across many representative workloads.