GLM

by Zhipu AI · Website

Zhipu AI's GLM family of flagship reasoning models — roughly 750B parameter MoE architectures with 40B active parameters per token, released under MIT license. GLM-5.1 advanced long-horizon autonomous task capabilities, and GLM-5.2 (June 2026) added a usable 1M-token context with IndexShare sparse attention, posting frontier-class coding results. Their massive size means local use requires 512 GB-class hardware or aggressive quantization; most users run them via cloud tags.