GLM-5.2
Zhipu AI·
LLMsopen-weightcloud + localFrontier
Overview
753B Mixture-of-Experts (~40B active) open-weight model under MIT license. Introduces the IndexShare attention mechanism (~2.9x lower per-token FLOPs at 1M-token context) and a usable 1M-token context window. Ships with two thinking-effort levels; the highest tier is marketed as 'Max'. Available as open weights on Hugging Face and via the Z.ai API. Vendor-reported benchmarks (maximum thinking effort): GPQA Diamond 91.2, SWE-bench Pro 62.1, Terminal-Bench 2.1 81.0, HLE 40.5 (no tools).
Capabilities and innovations
753B total / ~40B active MoE1M-token contextTwo thinking-effort levels (incl. 'Max')Agentic / long-horizon codingMIT licenseIndexShare attention (~2.9x FLOP reduction at 1M-token context)Usable 1M-token context window
Benchmarks
| Benchmark | Score | Source |
|---|---|---|
| GPQA Diamond | 85.61% | vendor Z.ai self-report: GPQA Diamond 91,2%. Not used as primary value (deviation >3 points from the independent runs). |
| Humanity’s Last Exam (no tools) | 40.5% | vendor HLE 40.5 no-tools (recorded for cross-model comparability). Z.ai also reports 54.7 with tools, which is not comparable to the no-tools HLE used here. |
| SWE-bench Pro | 62.1% | vendor SWE-bench Pro 62.1 — Z.ai-reported (maximum thinking effort). |
| MMLU-Pro | 86.71% | unsourced |
| Terminal-Bench 2.x | 67.79% | independent Terminal-Bench 2.1 67.79%, measured by Vals AI with the Terminus 2 harness (accessed 2026-09-23). Vals AI is the reference harness for this field so values compare across vendors; vendor and other third-party figures stay in the notes below. Replaces 77.9 (vendor / third_party). |
| Terminal-Bench 3.0 | 4.6% | independent Terminal-Bench 3.0 4.6% (± 1.9, 95% CI), official leaderboard maintained by the Terminal-Bench team (Stanford / Harbor / Laude Institute), max effort, agent harness: Claude Code (accessed 2026-09-23). Matches the value Z.ai reported in the GLM-5.3 launch comparison. Do not compare against the terminal_bench column: 3.0 is a harder generation on a different scale. |
| DeepSWE v1.1 | 46.2% | vendor DeepSWE v1.1 46.2 — Z.ai's GLM-5.3 launch table reporting GLM-5.2 as its own predecessor baseline (GLM-5.3: 66.9). Vendor self-report over its own model. Datacurve's own DeepSWE board was not reachable from this environment, so the value is not cross-checked against it. |
- MMMU-Pro: not reported by the vendor
Architecture and hardware
- Parameters
- 753B total, 40B active per token (MoE)
- Estimated VRAM at Q4
- ~433 GB, Frontier class
Links
More from Zhipu AI
Data curated by AI Model Timeline. See the methodology for admission criteria, benchmark eras and source priority.