A Chinese AI company called Zhipu released a new model called GLM-5.2 on June 16, 2026. Think of it as a large assistant specialized in reading and writing code — and it comes free to use and modify. (z.ai)

The key feature is its memory size. Most AI models can hold about as much information as a small book chapter in their "mind" at once. GLM-5.2 can hold the equivalent of an entire shelf of books — enough to read your entire codebase or a company's full financial records without forgetting what it started reading. For teams that automate code review or write tests automatically, this is useful because they do not need to keep fetching and reorganizing information.

The model also has two thinking speeds. For simple tasks like suggesting a line of code, it can work quickly. For harder questions, it can think more carefully. This lets teams choose what speed is worth the cost for each job, rather than running everything at maximum intensity.

If you already use Zhipu's coding tool, you get access to GLM-5.2 automatically.

The model can also check facts across different sources. If a financial news article mentions a company's profit figure, the model can verify that number by looking at the official company filings and private investment records. It can cross-check information across multiple sources in sequence — something that used to require either a much smarter AI or a person manually checking each source.

Here is where the money side matters. Most powerful AI models charge companies for every word they process. When a company runs thousands of automated code reviews or tests, the costs add up fast. But GLM-5.2 is released in a way that lets teams run it on their own computers instead of paying per use. They pay only the electricity and computer power to run it themselves. For large teams, that difference is real.

The model also comes with a legal license that lets companies use it for business without any restrictions. Other free AI models sometimes come with rules that confuse lawyers and slow down adoption in companies. GLM-5.2 avoids that problem.

The ability to switch between thinking speeds reflects a shift in how companies build AI workflows. Not every decision needs the slowest, most careful thinking. A quick suggestion inside a larger automated system is different from a final review. Letting teams adjust the effort level means they can keep costs down while still being careful where it matters.

Zhipu is betting on two markets here. First, software development — code review, test writing, and documentation are areas where AI can help right now. Second, knowledge work like investment research and compliance, which need the same kind of careful, multi-step verification that coding does.

GLM-5.2 is available now through Zhipu's coding tool and as free, downloadable weights from HuggingFace and ModelScope.