On February 11, 2026, Zhipu AI released GLM-5 under an MIT license — 744 billion parameters with a Mixture-of-Experts architecture, rivalling Claude Opus 4.5 and GPT-5 on standard benchmarks. Without a single NVIDIA GPU in its training stack.
The Architecture: MoE at Scale
744B Total Parameters
Only ~180B activate per forward pass — making inference economically viable at scale.
256 Expert Modules
128 experts activate per token, routing by a learned gating network.
128K Context Window
Competitive context length for long-document analysis and agentic tasks.
Multilingual by Design
Strong in Chinese, English, Japanese, and Korean — native multi-language training.
Benchmark Performance
| Benchmark | GLM-5 | Claude Opus 4.5 | GPT-5 |
|---|---|---|---|
| MMLU (5-shot) | 91.2% | 92.1% | 91.8% |
| HumanEval | 88.4% | 89.0% | 90.2% |
| MATH | 79.6% | 81.3% | 82.1% |
| C-Eval (Chinese) | 94.3% | 88.7% | 87.9% |
Context: US export controls restricted NVIDIA's most powerful chips from Chinese buyers since 2022. Rather than limiting Chinese AI, these controls appear to have accelerated domestic alternatives.
An MIT-licensed frontier model from China changes the calculus for every enterprise that was waiting for open-source quality to reach GPT-4 level. That threshold has now been crossed — and then some.
Our AI Automation team is evaluating GLM-5 for client deployments where data sovereignty requirements preclude cloud-hosted APIs. Contact us if you are assessing open-source model deployment.
Tags:
AI ModelsOpen SourceGLM-5ZhipuLLMChina AI
