GLM-5.2: A 1M Token Code Model, Mit License Coming Soon

On June 13, Z.ai announced the release of their latest model, GLM-5.2. This is a coding model with a 1M token context size and two effort settings (“High” and “Max”). This means that the model can be easily integrated into existing tools using an Anthropic compatible API, allowing you to swap it in without much work. The weights will be released under an MIT license (although this isn’t available right now, Z.ai claims it will be next week).

The most exciting feature here is the context size. With a 1M token context size, this is approximately 5x larger than GLM-5.1. And this time, when they say “usable”, they mean it — other models have advertised large context sizes that don’t actually work well above a couple hundred thousand tokens. In the case of coding, this means that the agent can remember all the code, tests, and documentation for a medium sized codebase at once, rather than having to make assumptions based on only a few files. Response length is capped at around 131,000 tokens, which should allow you to make significant changes to a codebase in a single interaction.

1What "High" and "Max" actually change

The two effort settings allow you to control how much thought goes into each response. “High” effort is the default, while “Max” effort uses more computational resources to solve complex tasks like tricky refactorings or bugs spread across multiple files. Of course, this comes at the expense of both time and tokens. This is part of a broader trend this year — Microsoft CEO Satya Nadella has cautioned against “token-maxing”, or using the highest possible effort setting for simple tasks. Learning to balance effort and task is important, rather than always using max effort.

GLM-5.2 is a 744 billion parameter Mixture-of-Experts model which uses 40 billion parameters per token. This may seem like a lot of parameters, but it’s relatively efficient because it only uses a small amount of parameters for each query. Unfortunately, Z.ai did not release any benchmark results when they released the model. There are no results for SWE-bench, Terminal-Bench, or even Code Arena. This is a problem, since we don’t have any objective data on how well the model performs until the weights and scores are released.

So why should non-developers care? The Anthropic compatible API is the real news here. We’re starting to see more and more tools follow this API format, which allows the tool and the underlying model to be separated. This means that if a model is taken down (like Anthropic’s Fable was last week) or if the pricing suddenly changes, you can switch to another model without having to change how you interact with the tool. And with open weights available under an MIT license, who knows — maybe you’ll be able to run it yourself someday!