GLM-5.2 is a large open-weights Mixture-of-Experts model from the Z.ai GLM family, with 753B total parameters and 40B active per token across roughly 1.51TB of weights distributed under an MIT license. It arrived in mid-June 2026, first through a Z.ai coding plan and then as a public open-weights release, and serves as the base architecture that the subsequent GLM-5.3 release inherits unchanged. Its design centers on text input and output, with very long context handling supported by an IndexShare-based long-context stack and reinforcement-learning infrastructure (SAO) plus asynchronous training (slime) that were built out in tandem with the model.
In practical use, GLM-5.2 targets reasoning-heavy and agent-style workloads including coding, tool calling, structured output, and long-horizon tasks, where its open weights give teams flexibility to self-host. Independent testing on the Artificial Analysis Intelligence Index notes that it is more token-hungry than peers such as GLM-5.1 or MiniMax-M3, producing around 43k output tokens per task, which is an important consideration for inference budgeting. The same base model was later pushed further by post-training alone to produce GLM-5.3, which adds large gains on coding benchmarks, long-horizon agent tasks, and emergent cyber capability work, signalling where the GLM line is headed next.
