753B sparse MoE optimized for long-horizon autonomous coding. 1M-token context via IndexShare sparse attention (2.9× compute reduction). MIT license on Hugging Face.
Released June 2026; surpasses GPT-5.5 on engineering benchmarks at 1/6 the cost.
Updated 13 Jun 2026Checked 10 Oct0 updates this week
Dynamic token selection and sparse patterns achieve 2.9× compute reduction.
Trained on diverse agent trajectories: multi-step coding, planning, long-context reasoning.
Reason over images, tables, and code simultaneously.
| Spec | GLM-5.2 |
|---|---|
| Total parameters | 753 billionR (reported) |
| Active parameters per token | ~40 billionR (reported) |
| Context window | 1 million tokensR (reported) |
| Terminal-Bench 2.1 score | 81.0R (reported) |
| SWE-bench Pro score | 62.1R (reported) |
| Compute reduction at 1M context vs dense | 2.9× via sparse attentionR (reported) |
| Competitive advantage vs GPT-5.5 | ~1/6 cost for same capabilityR (reported) |
R reported by the company
Z.ai (Zhipu AI)
GLM-5, GLM-5.2, open-weight models, coding agents
China's first major LLM company to go public. Founded 2019 by Tsinghua professors. Develops GLM open-weight models. Debuted Hong Kong Stock Exchange January 2026 at $7.4B, raising $558M.
