GLM 5.1
zai-org · released 2026-04-07 · mit license
GLM-5.1 is Z.ai's open-weight flagship for agentic engineering (April 2026), leading SWE-bench Pro at 58.4% at launch. Strongest on coding and agentic tasks; trails the frontier on pure reasoning.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | 753.86B total · MoE, — active |
| Architecture | glm_moe_dsa |
| Context window | 203K tokens |
| Knowledge cutoff | — |
| Modalities | text |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 86.2% |
|---|---|
| SWE-bench Verified | 77.8% |
| AIME | 95.3% |
| MMLU-Pro | — |
| BFCL v3 (tool use) | — |
| Composite score | 7.32 |
| Community rating | No reviews yet |
VRAM & disk per quantization
| Quant | VRAM | Disk | RAM | Context |
|---|---|---|---|---|
| Q4_K_M | 438.7 GB | 437.2 GB | — | 203K |
Strengths & weaknesses
Strengths: Leads SWE-bench Pro (58.4) among open weights at release; Strong long-horizon agentic coding and tool use; Open weights under permissive MIT
Weaknesses: GPQA (86.2) and HLE (31.0) trail the top closed models; 200K context smaller than DeepSeek V4's 1M