GLM-4.5
Open weightsby Zhipu AI·China·Released
Zhipu's flagship — agentic-first MoE with strong coding + tool-use benchmarks.
About this model
GLM-4.5 (July 2025) is Zhipu AI's flagship — a 355B-parameter MoE (32B active) released under MIT. Zhipu (founded by Tsinghua University researchers) positions GLM-4.5 explicitly as the 'agentic-first' open-weights model, with strong tool-use and function-calling benchmarks alongside competitive general reasoning.
MIT licensing is unusually permissive — even more so than Apache 2.0 in some interpretations — and combined with the model's BFCL v3 score of 77.8% (a tool-use benchmark) makes GLM-4.5 particularly suited for agent frameworks. The model is served via Zhipu's z.ai API at aggressive pricing and also available as open weights on Hugging Face.
Strengths
- •MIT licensed — most permissive of any frontier-adjacent model
- •Industry-leading tool-use / function-calling benchmarks (BFCL v3 77.8%)
- •Strong general reasoning (MMLU-Pro 84.6%)
- •Active 32B parameters — runs on smaller serving infra than DeepSeek V3
Limitations
- •English chat quality trails Western frontier models
- •Smaller research footprint than Qwen or DeepSeek
- •Hosted API less battle-tested for production international workloads
When to use it
- →Agent frameworks needing reliable tool calling
- →Open-weights workloads under strict licensing requirements
- →Chinese-market enterprise deployments with research-lab backing
Architecture & training
Zhipu's GLM-4.5 paper (arXiv 2508.06471) documents a hybrid SFT + RL post-training pipeline with specific emphasis on tool-use traces. The 355B/32B-active configuration places it between DeepSeek V3 and Llama 4 Scout in scale.
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| BFCL v3 | 77.8 | |
| MMLU-Pro | 84.6 | |
| SWE-bench Verified | 64.2 |