TL;DR
An official release of the 750B-A40B GLM-5.2 model, combining a one-million-token context, IndexShare sparse-attention reuse, and larger-scale agentic reinforcement learning for long-horizon tasks.
Why it matters
GLM-5.2 links context scaling to RL systems design. Longer trajectories create harder scheduling, compaction, verification, and reward-hacking problems, so the release is useful evidence that context length and environment throughput have to scale together.
Key findings
- 01
IndexShare reuses one indexer across groups of sparse-attention layers; Z.ai reports 2.9× lower per-token FLOPs at one-million-token context.
- 02
The agentic RL program expands the scale, domain coverage, and execution complexity of training tasks while supporting white-box, black-box, compact-trajectory, and sub-agent rollouts.
- 03
The team reports using slime for large-scale agent RL and parallel online policy distillation, while adding anti-hacking checks for verifiable coding rewards.
Scaling dimensions
Models, methods & benchmarks
- Models
- GLM-5.2, GLM-5.1
- Algorithms
- SAO, IndexShare, Online Policy Distillation, slime
- Benchmarks
- Terminal-Bench 2.1, FrontierSWE, PostTrainBench, SWE-Marathon
Topics