Chinese AI which runs locally challenges premium models like ChatGPT and Claude
&imwidth=600&imheight=450&format=webp&quality=medium "Chinese AI which runs locally challenges premium models like ChatGPT and Claude")
Chinese AI company releases GLM 5.2, an open model that runs locally.
Chinese AI company called Z.ai (formerly Zhipu AI) has released GLM 5.2, which challenges the idea that you need to pay premium subscription prices to big companies for frontier-level performance. However, open models have their problems.
A new open-weight AI model is fast becoming popular, challenging expensive, premium subscriptions like ChatGPT or Claude. A Chinese AI company called Z.ai (formerly Zhipu AI) has released GLM 5.2, which offers frontier-level performance that can be run entirely on local hardware. It goes against a major assumption - that premium subscriptions are the only way to go. ChatGPT and Claude are closed systems, while GLM 5.2 gives developers direct access to the model itself.
Open Source gives you everything, from model weights to training code, a data processing pipeline, an evaluation framework and often the training dataset. Open-source and open-weight AI models are highly flexible and cheaper, but earlier fell behind bigger models. But Meta’s Llama family, Mistral, and GLM 5.2, are now highlighting that not every task needs the most expensive AI model. Not every business needs to solve world-class theoretical logic, and simply needs open models that can complete the majority of their work at a low cost.
Challenges with open models
GLM 5.2 is a Mixture-of-Experts (MoE) model boasting 744-billion to 753-billion parameters. It features a 1-million-token context window, meaning it can process entire codebases or libraries of books simultaneously. Being an "open-weight" model, it lets developers download its "brain" directly. This allows companies in sensitive industries, like finance or medicine, to keep all data in-house instead of sending it to third-party corporate cloud servers.
However, the open models have severe hardware requirements. Running GLM 5.2 locally comes with intense physical limitations. In its uncompressed form, it takes up a staggering 1.51 terabytes of storage and memory. It needs to be heavily compressed, even if it needs to be loaded on consumer-grade hardware. Even in this case, it demands roughly 240GB of memory just to load. GLM 5.2 might not replace everything, but it shows that open models are becoming cheaper and more competitive.
Trending Stories
&im=FitAndFill=(700,400))
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
)
)
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
)
)
&im=FitAndFill=(700,400))
)
)
)
