GLM-5.2

GLM-5.2 is presented as a flagship release focused on long-horizon tasks, meaning extended multi-step coding and agent workflows rather than single-turn answers. The defining capability is a solid one-million-token context designed to stay coherent across very long, messy agent trajectories, framing the model as a foundation for engineering-heavy assistants that must keep state and reasoning intact across many turns and tool calls.

The model introduces architectural changes aimed at making that long context practical rather than just nominally large. It uses an IndexShare sparse attention design that reuses a single indexer across every four sparse attention layers, reportedly cutting per-token compute by about 2.9× at 1M tokens, and upgrades the MTP speculative-decoding layer to lift acceptance length by up to 20%. Coding capability is described as notably stronger than the prior GLM-5.1, with configurable thinking-effort levels to trade latency for depth. The release is distributed as a fully open weights model under an MIT license with no regional access limits, suiting teams that want on-premise or self-hosted long-context coding agents without proprietary restrictions.