Ant Group and vLLM open-source Humming inference library

Most of the time a GPU spends answering a prompt goes into multiplying large grids of numbers. In October, Ant Group and the vLLM community released Humming, an open-source library that generates the code for those multiplications on demand, tuned to the precision, model layout and chip in use. It is published under the Apache 2.0 licence from Ant Group's inclusionAI organisation and mirrored under the vLLM project.

Quantization stores a model's weights with fewer bits: INT4 uses 4 bits per weight where FP16 uses 16, which shrinks memory use. Each combination of weight width, activation format, scaling scheme, dense or mixture-of-experts layout, and GPU generation has tended to need its own kernel, the low-level code that runs on the chip. Humming converts supported formats into one shared description and compiles kernels from reusable CUDA components. It covers integer weights from INT1 to INT8 and floating-point weights from FP4 to FP8, on dense and MoE (mixture-of-experts) layers, and targets NVIDIA GPUs from Turing onward. Its dependencies are limited to PyTorch and NVCC, NVIDIA's compiler, and the pip package is about 100 KB.

The release also points beyond NVIDIA. It says Humming already fully supports T-Head's Zhenwu 810E, a CUDA-compatible AI accelerator from Alibaba's chip unit, and that the same kernel and tuning path can run on that device. Zhenwu 810E has already appeared in earlier Chinese AI serving deployments. The performance figures come from the teams' own microbenchmarks and serving tests and have not been independently verified. At small batch sizes, dense W4A16 (4-bit weights, 16-bit activations) is reported at roughly 3 to 4 times cuBLAS on several devices, and on B300, W4A8 reaches about 1.6 times a DeepGEMM baseline in some small-to-medium batch cases.

The most concrete test is end-to-end. On eight H20 GPUs, running DeepSeek-V4.1-Flash with the MoE layers switched from Marlin W4A16 to Humming W4A8 raised output throughput by about 14% to 43% and cut mean time to first token, the wait before the first word appears, by about 25% to 32%. The largest gains came at higher concurrency, that is, when more requests are served at once.

For operators, the change is a configuration switch rather than a rewrite. From vLLM 0.29.0, CUDA installs of vLLM include Humming as a dependency, selectable with --linear-backend humming and --moe-backend humming. A team serving quantized models on NVIDIA hardware can run its own traffic with those flags, and that measurement will settle how much of the reported range applies to its workload. The Zhenwu support is what the release states; the broader speed figures still await outside benchmarks. Humming is a joint Ant Group and vLLM project, not an Ant-only product.