The model runs on 753 billion parameters and holds a full one million tokens of context, which basically means it can read an entire codebase and not lose track of what it just did. Z.ai released the whole thing under an MIT license on Hugging Face, so anyone can download it and run it on their own machines instead of sending code to a server overseas. And within days of the coding plan going live, the company opened up standalone API access too. Pricing for the lightest tier starts at $12.60 a month, which is less than a lot of people pay for a single streaming service, and that number alone has developers talking.
On Arena's Code Arena leaderboard, GLM 5.2 landed in second place overall, trailing only the newest Claude model and sitting ahead of everything else currently available for public use. There's chatter about Huawei hardware being involved in training it, though nobody's confirmed that yet, and if it's true it says a lot about how far Chinese chipmakers have come. Something about a model this capable costing less than a gym membership feels almost unfair to the labs that spent billions getting here. Developers are already comparing it to the DeepSeek moment from last year, except this time there wasn't nearly as much panic.
Comments