The company calls the approach W4A16, contrasting it with conventional W4A4 paths that quantize activations during decode.
According to the October 23, 2026 post, the kernel removes activation quantization from the critical path and overlaps weight dequantization with computation and data movement. It uses Blackwell features including native NVFP4 conversion instructions, tensor memory, asynchronous execution, and Tensor Memory Accelerator and matrix-multiply-accumulate instruction handling. humans& says its implementation outperformed FlashInfer's optimized W4A4 MegaMoE path by up to 25% in a GLM-5-shaped layer test across eight B300 GPUs and 600 samples per configuration.
The source presents a company benchmark, not an independent evaluation. It does not establish end-to-end serving gains for a named production model, performance on GPU generations other than Blackwell, cost changes, or general accuracy results across workloads.