A Chinese AI lab called Zhipu AI has released a new artificial intelligence model designed specifically to help with computer code. The model, called GLM-5.2, says it can look at one million tokens—essentially one million small pieces of text or code—all at once. To put that in perspective, most coding models today can only look at 128,000 to 256,000 tokens. That is roughly the difference between reading a single long chapter versus reading an entire book.

But there is a catch. A larger view does not automatically mean a better view. Think of it like memory: you can have a big backpack, but if you cannot find what you need inside it quickly, the extra space does not help. Zhipu says its model does not lose accuracy when working with all one million tokens—that is the "lossless" part of the claim. Whether that holds true in practice will become clear once researchers outside the lab run their own tests (Zhipu AI).

There is another claim that may matter more. Zhipu says the model stays focused and consistent when asked to perform long chains of edits and tasks—for example, rewriting code across an entire project and then testing it. Most AI coding tools today fail not because they run out of room to think, but because they get confused and inconsistent after many steps. A model that does not lose its way through 50 or 100 sequential steps is the difference between a tool that shows promise and one you can actually use at work.

Zhipu is one of several Chinese AI labs competing to release powerful models that developers can use freely. Similar labs like DeepSeek and Moonshot have done the same thing, treating open releases as a way to attract engineers who might otherwise use models from bigger American companies like Google or OpenAI.

Zhipu has not published detailed test results showing exactly how well GLM-5.2 performs compared to other models. That means the claim of being the "best open-source coding model" is not independently verified yet. Additionally, there is a difference between two kinds of releases: in one, the lab shares the model weights (the core AI) but keeps the training process and data private; in the other, the lab shares everything, including how it was built. For companies thinking about using this model, the legal terms and what you are allowed to do with it will matter a lot.

The timing here is worth noting. Long context windows and the ability to do long chains of work are becoming the main battleground between AI companies right now. Google, Anthropic, and OpenAI have all been competing to make better long-context models. Open models from labs like Zhipu are catching up to those expensive proprietary services. If GLM-5.2 truly works as described, it means companies no longer have to pay per-token to big AI providers for coding help—they can use an open model on their own computers.

For anyone considering GLM-5.2, the practical questions are simple: Is it legal to use for your business and to modify? Does it actually work well at the sizes your project needs? Does it handle the kind of multi-step work your team actually does? The same checks apply no matter who builds the model. The advantage of an open model is that you can run these checks yourself instead of just trusting what the company says.