Chinese AI lab Zhipu AI has released GLM-5, a 744-billion parameter mixture-of-experts (MoE) model, under the MIT license. The model uses 44 billion active parameters per forward pass, features a 200,000-token context window, and scores an impressive 77.8% on SWE-bench Verified.
Model Specifications
| Spec | Value |
|---|---|
| Total parameters | 744B |
| Active parameters | 44B (MoE) |
| Context window | 200K tokens |
| SWE-bench Verified | 77.8% |
| Training hardware | Huawei Ascend |
| License | MIT |
Why This Matters
Open-Source Under MIT
The MIT license is the most permissive available — anyone can use, modify, and commercialize the model without restrictions. This is a stronger commitment to open-source than Meta's Llama license, which includes usage-based restrictions for large companies.
Trained on Huawei Ascend
GLM-5 was trained entirely on Huawei's Ascend AI chips rather than NVIDIA GPUs. This is significant because US export controls have restricted China's access to advanced NVIDIA hardware. GLM-5's strong benchmark performance demonstrates that the Chinese AI ecosystem can produce competitive models despite these restrictions.
SWE-bench Performance
The 77.8% score on SWE-bench Verified — a benchmark that tests a model's ability to solve real-world software engineering tasks from GitHub issues — places GLM-5 among the top-performing models on this widely-watched benchmark, alongside tools like Moonshot's Kimi Code which scored 62% on the same test.
The Chinese Open-Source Wave
GLM-5 is the latest in a series of strong open-source releases from Chinese AI labs:
- DeepSeek has released several competitive open models
- Alibaba's Qwen series continues to improve
- Baichuan and 01.AI have contributed significant open-weight models
This trend is creating a two-track open-source ecosystem, with both Western labs (Meta, Mistral) and Chinese labs (Zhipu, DeepSeek, Alibaba) producing frontier-quality open models. Alibaba's Qwen3.5, for example, offers 201-language support with an agent-first architecture.
Availability
The model weights, tokenizer, and training documentation are available on Hugging Face and ModelScope. Zhipu has also published a technical report detailing the architecture, training process, and evaluation methodology.
Community members have already begun creating quantized versions for consumer hardware, though the full model requires significant compute resources to run.
