Understands text and images, writes code, and runs multi-step tasks

You give it a question, a screenshot, or a full coding task, and it answers, explains what it sees, or writes the code. It is the first model in the new GLM-5 family built to be multimodal from the ground up: the same model that fixes your Python script can also read the error dialog you screenshotted.

The practical difference is speed and cost. Z.ai built it as a mixture-of-experts model: it stores 320 billion parameters of knowledge but activates only 18 billion of them for any given word, so it responds like a much smaller model. The company reports it comes close to Claude Opus 4.8 on coding and agent-style performance tests, and beats its own predecessor at a tenth of the price.

The weights are on Hugging Face under the MIT license, which means you can use them in a commercial product without restrictions.

There is no browser demo. The quickest route is Z.ai's hosted API: paste in a screenshot of an error dialog and ask for the fix, or hand it a small refactoring task, and compare the answers to the assistant you use today. If your team runs its own models, the model page links ready-made vLLM and SGLang recipes, but be realistic about scale: the full weights need server-class GPUs, not a desktop card. For most readers the hosted API is the honest first test.

The release is days old, so the coding and agent numbers are the vendor's own; independent results will take a few weeks. And while the license allows self-hosting, the model's size makes that a real infrastructure project rather than a weekend install.

• Paste an error screenshot and ask what went wrong and how to fix it.
• Hand it a coding task inside an agent tool and let it work through the steps.
• Ask questions about a long report, charts included, in one conversation.
• Prototype on the hosted API now, move to your own servers later without switching models.

Try the demo

View model page

Read this issue on neuralexpedition.com →