Two days testing GLM-5.2: the good and the bad

I spent two days trying GLM-5.2. This thing can do heavy lifting, but don't expect it to do everything.

First, how to get it running. GLM-5.2 is Zhipu AI's flagship model, with weights open-sourced in June this year. Open weights are attractive for people doing research — it means you can download the model files and deploy them yourself, with data never leaving the intranet. I didn't build my own cluster; I went through a domestic cloud platform's managed entry point. Registration, real-name verification, creating an API key — less than twenty minutes total. Let me insert an old saying here, which I also mentioned last month when writing about another model: don't hardcode your keys in scripts, put them in environment variables first. Newbies trip up on this step nine times out of ten, and it has nothing to do with the model itself.

First thing was testing whether it can honestly spit out structured data. Structured means having it output in a fixed format, like JSON, so programs can process it further. I threw in a log with dates and asked it to extract by field. The result was cleaner than I expected — it consolidated two columns of inconsistently formatted dates on its own, instead of giving me a prose paragraph like some models do. High marks for this one.

Then long context. This model's selling point is 1M tokens. A token can be roughly understood as the unit of word count in the model's eyes; 1M is about a thick book, or the code volume of a medium-sized project. I stuffed an entire small project I maintain into it and asked it to explain how the modules call each other. It could answer, and even pointed out two circular dependencies I hadn't noticed myself. This step really has something to it.

Where it gets stuck also needs mentioning

First sticking point: it doesn't generate images. This is a pure text model. If you ask it to draw a flowchart, it can only give you SVG code or ASCII art, which you have to render yourself. I didn't catch on at first, thought my prompt was wrong, and tried three rounds before confirming — it just can't. If you want to do integrated text-and-image work, switch tools early.

Second is latency. The fuller the context, the longer the wait. Long-horizon tasks sound nice, but in practice you're trading time for it. From my testing, once it's full, a single reply takes quite a while. Exactly how long depends on task complexity — hard to generalize.

Third isn't its fault, but something I found while checking sources. Two sources disagree on the launch date — one says June 17, the other differs slightly. This kind of small discrepancy is pretty common in domestic model press releases. Not fatal, but a reminder: when you see leaderboards and dates, verify from multiple sources before drawing conclusions.

One more point. Some evaluations say its scores on several long-horizon task test sets fall between Claude Opus 4.7 and 4.8, with one item only about 1 percentage point below 4.8, making it the highest-ranked open-source model.

Long-horizon task test sets look at whether AI can, like a top engineer, complete a complex project on a scale of hours to tens of hours.

I can't reproduce this data myself, so I'll just note it down and not use it as a conclusion.

Who it's for, who it's not

Suitable for developers with large codebases who want to feed everything in at once for analysis, and also for teams building agents. Agents are programs that can call tools and work step by step on their own; its tool calling and MCP support are complete. Teams with data compliance requirements who want to deploy open weights themselves can also consider it.

Not suitable for image generation, not for people who just want to chat with it cheaply and quickly. And definitely not for people who don't understand key management and API call flows but want to wrap it and launch — I've seen too many of this last type, and problems always happen at this step.

One action recommendation: don't stuff an entire project in right away. Start with one file, one function, get a single call working, confirm the key, billing, and return format are all correct, then scale up. Long context is both a capability and a cost. Before you start, think clearly: do you really have to stuff that much in?