GLM 5.2 is positioned as a flagship release focused on long-horizon work rather than short exchanges. The defining step is the move to what the release describes as a "solid 1M-token context" that can sustain extended, messy coding-agent trajectories in practice, not merely accept more tokens. To make that long window affordable, the team introduced an architectural refinement called IndexShare, which reuses a single indexer across every four sparse attention layers and reduces per-token FLOPs by roughly 2.9× at a 1M context length. The release also highlights coding as a primary capability, with multiple "thinking effort" levels that let callers trade depth of reasoning against latency on agentic coding workflows.
GLM 5.2's broader contribution is a reusable training stack designed to keep improving on the same base model. Alongside IndexShare, the team built SAO for reinforcement learning on long-horizon tasks and the open-source slime framework for large-scale asynchronous training, both running on accumulated long-horizon task environments. The MIT-licensed, open-weights distribution makes the model attractive for teams building autonomous coding agents, research assistants, or any pipeline that needs to stay coherent over very long prompts. A subsequent release built on top of GLM 5.2 has already demonstrated large gains in complex coding and emergent cyber capability by continuing to scale this same stack, reinforcing GLM 5.2's role as the foundation model for an ongoing long-horizon research line.
