GLM 4.5
zai-org · released 2025-07-25 · mit license
GLM-4.5 is Z.ai's 355B-parameter (32B active) MoE (July 2025, MIT); per its technical report it scores 64.2% SWE-bench Verified, 91.0% AIME 24, 79.1% GPQA and 84.6% MMLU-Pro. A top-3 open model at release.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 358.34B total · MoE, — active |
| Architecture | glm4_moe |
| Context window | 131K tokens |
| Knowledge cutoff | 2024-12-31 |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 79.1% |
|---|---|
| SWE-bench Verified | 64.2% |
| AIME | 91% |
| MMLU-Pro | 84.6% |
| BFCL v3 (tool use) | — |
| Composite score | 6.57 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 209.3 GB | 207.8 GB | — | 131K |
Strengths & weaknesses
Strengths: Strong agentic/tool-use on par with Claude Sonnet 4 at release; Excellent math (AIME 91.0, MATH-500 98.2); MIT license
Weaknesses: Coding (SWE-bench 64.2) below top proprietary of its era; Superseded by GLM-4.6 and GLM-5