The listed Chinese AI company Zhipu AI has introduced GLM-5, the fifth generation of its large language model. The model comprises 745 billion parameters and is expected to be released as open source under the permissive MIT license. Particularly noteworthy: GLM-5 was trained entirely on Huawei Ascend chips without the use of NVIDIA hardware.
What is GLM-5?
GLM-5 is a large language model developed by Zhipu AI, a company founded in 2019 as a spinoff from Tsinghua University. The model uses a Mixture-of-Experts architecture (MoE) with a total of 745 billion parameters. Of these, 44 billion parameters are active per inference operation. The system features 256 experts, of which 8 are activated per token.
The model was developed for complex tasks in the areas of programming, logical reasoning, and autonomous agent systems. It can process context windows of up to 200,000 tokens and uses the DeepSeek Sparse Attention mechanism for efficient processing of long sequences. GLM-5 is the successor to GLM-4.5, which comprised 355 billion parameters (32 billion active).
| Specification | Value |
|---|---|
| Architecture | Mixture of Experts (MoE) |
| Total Parameters | approx. 745 billion |
| Active Parameters | approx. 44 billion |
| Expert Configuration | 256 total / 8 active (5.9%) |
| Context Window | up to 200,000 tokens |
| Training Hardware | Huawei Ascend |
| Framework | MindSpore |
| License | MIT (expected) |
Open Source with MIT License
Zhipu AI has a clear tradition of open-source releases. The current flagship model GLM-4.7 is already freely available on Hugging Face for commercial use. For GLM-5, it is expected to follow this example and be released under the MIT license. This license permits unrestricted commercial use, adaptation, and further development.
The model weights of GLM-5 are already publicly accessible on Hugging Face and ModelScope. The model supports various inference frameworks such as vLLM and SGLang for local implementations. In addition to open-source availability, GLM-5 is also available as an API through the developer platforms api.z.ai and BigModel.cn.
Pricing remains an important advantage of the GLM series. API prices for GLM-4.x are approximately 0.11 US dollars per million tokens, significantly cheaper than comparable models such as GPT-5 (1.25 US dollars per million input tokens, 10 US dollars per million output tokens). For GLM-5, this price advantage is expected to be maintained or improved.
Training Without NVIDIA Chips
A strategically significant aspect of GLM-5 is complete independence from US hardware. The model was trained exclusively on Huawei Ascend chips using the MindSpore framework. This represents a milestone in China’s effort to build an independent AI infrastructure.
Financing for this development was enabled in part by an initial public offering in Hong Kong on January 8, 2026, in which Zhipu AI raised approximately 4.35 billion HKD (558 million US dollars). These funds directly accelerated the development of GLM-5 and position the company for further investments in next-generation AI architectures.
Zhipu AI also supports the implementation of GLM-5 on other non-NVIDIA chips, including Moore Threads, Cambricon, Kunlun Chip, MetaX, Enflame, and Hygon. Through kernel optimization and model quantization, GLM-5 can achieve appropriate throughput on these platforms as well.
Performance and Availability
According to Zhipu AI, GLM-5 achieves top performance among all open-source models in the areas of logical reasoning, programming, and agent systems. In benchmarks such as Humanity’s Last Exam (with tools: 50.4 points), SWE-bench Verified (77.8%), and Terminal-Bench 2.0 (56.2% and 60.7% in the verified version), the model demonstrates competitive results compared to proprietary models such as Claude Opus 4.5 and GPT-5.
Particularly noteworthy is the performance on long-term tasks. In Vending Bench 2, a benchmark for long-term operational capabilities, GLM-5 achieved an account balance of 4,432 US dollars, ranking it in first place among all open-source models.
GLM-5 is accessible through multiple channels. Users can use the model through the Z.ai chat platform in chat mode or agent mode, access it via APIs, or implement it locally with the publicly available model weights. For developers, Zhipu AI also offers integrations with coding agents such as Claude Code, Cline, and others through the GLM Coding Plan.
With GLM-5, Zhipu AI positions itself as a serious competitor in the global AI market and demonstrates that top-tier models can be developed outside of established US AI infrastructure.
How good and how widespread GLM-5 ultimately becomes remains to be seen; however, in the ranking from Artificial Analysis, which subjects LLMs to a series of tests, one can already see that the new AI model under MIT license is currently the strongest open-source model in the world:
