GLM 4.5 Air
zai-org · released 2025-07-25 · mit license
GLM-4.5-Air is the compact 106B MoE sibling of GLM-4.5 (July 2025, MIT); per its report it scores 57.6% SWE-bench Verified, 89.4% AIME 24, 75.0% GPQA and 81.4% MMLU-Pro, trading capability for efficiency.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 110.47B 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 | 75% |
|---|---|
| SWE-bench Verified | 57.6% |
| AIME | 89.4% |
| MMLU-Pro | 81.4% |
| BFCL v3 (tool use) | — |
| Composite score | 6.14 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 65.6 GB | 64.1 GB | — | 131K |
Strengths & weaknesses
Strengths: Compact 106B MoE with near-flagship reasoning (AIME 89.4, MATH-500 98.1); Efficient agentic/tool-use; MIT license
Weaknesses: Lower coding than full GLM-4.5 (SWE-bench 57.6); Superseded by newer GLM generations