GLM-5.1
Post-trained GLM-5 that tops SWE-Bench Pro at 58.4 and can work autonomously on one task for up to eight hours.
Same 744B MoE base as GLM-5, MIT-licensed. Z.ai demos include a 655-iteration run building a Linux desktop and a 6.9x vector-database throughput gain; it claims parity with Claude Opus 4.6 across 12 benchmarks.
- Date
- Tuesday, 7 April 2026
- Lab
- Zhipu AI / Z.ai
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-Bench Pro | 58.4 vs GPT-5.4 57.7 and Claude Opus 4.6 57.3 per press | company |
| Terminal-Bench 2.0 | 63.5 | company |
| CyberGym | 68.7 | company |
Weights appeared on HF 2026-04-03 UTC by repo creation, public launch 2026-04-07 per Z.ai release notes; Wikipedia says subscribers got it in March. Benchmarks are Z.ai-reported.
Sources
- huggingface.co/zai-org/GLM-5.1
- docs.z.ai/guides/llm/glm-5.1
- rits.shanghai.nyu.edu/ai/glm-5-1-z-ais-open-weight-model-takes-1-on-swe-bench-pro
This record was checked against its sources on 6 October 2026. How we check