Research & Innovation
Pruning 64 Experts Per Layer Cut a 306 GiB Model to 141 GiB With Benchmarks Inside the Noise
autotrust/GLM5.3-Flash-E224-DGX-Spark appeared on Hugging Face on October 6, 2026. It retains 224 of GLM-5.3-Flash's 288 routed experts, selected by neural architecture search, quantises them to NVFP4, and lands at 141.5 GiB — sized for a pair of ConnectX-7-linked DGX Sparks rather than for a datacentre. MIT licensed.