GLM Glossary
A quick reference of technical terms around GLM models: architecture, attention, benchmarks and product concepts.
| MoE Mixture of Experts | Each token activates only part of the expert networks, balancing model capacity against inference cost. |
| Post-training Post-training | The large-scale training stage after the pre-trained base (SFT/RL etc.); all GLM-5.3 gains come from here. |
| 1M Context Million-token context | A one-million-token context window that can ingest a large codebase or long document at once. |
| IndexShare Index sharing | One lightweight indexer shared by every four sparse-attention layers, cutting per-token FLOPs by 2.9× at 1M context. |
| KVShare KV sharing | Sharing key-value caches across layers to further compress long-context memory usage. |
| MTP Multi-Token Prediction | Predicting several future tokens at once, speeding up inference and training signals. |
| Thinking Level Thinking level | Three controllable reasoning depths — low / high / max — allocating compute by task complexity. |
| Terminal-Bench Terminal-Bench | Evaluates long-horizon system tasks in a real terminal; GLM-5.3 scores 88.2 on v2.1. |
| CyberGym CyberGym | Offense/defense security evaluation where GLM-5.3 leads at 84.5%. |
| ExploitBench ExploitBench | Source-code vulnerability audit and exploitation; GLM-5.3 scores 54.4 (previous gen 24.4). |
| Agent Agent | An AI system that plans autonomously, calls tools and completes multi-step tasks. |
| MIT License MIT License | A permissive open-source license: commercial use, modification and redistribution allowed with attribution. |