China's Z.ai Gives Away a Top-Tier Coding Model for Free

While the American labs trade blows over billion-dollar valuations and government access lists, a Chinese company has been quietly doing something that scares them more than any benchmark: giving a top-tier coding model away for free.

Z.ai has launched ZCode, a free desktop coding tool for Mac, Windows and Linux built around its open-weights model GLM-5.2. The model itself arrived in mid-June under an MIT license, meaning anyone can download the weights from Hugging Face, run them on their own machines and fine-tune them without asking permission or paying a cent.

Good enough to matter

This would be a footnote if the model were mediocre. It is not. GLM-5.2 is a 753-billion-parameter system that, on several long-horizon coding benchmarks, beats OpenAI's GPT-5.5, scoring 62.1 to GPT-5.5's 58.6 on the industry's SWE-bench Pro test, according to reporting from VentureBeat, and comes close to Anthropic's flagship Opus 4.8 on others. Independent reviewers have found it strong enough that it was briefly mistaken for a much larger Western model.

And it is cheap. Where it is offered as a paid service, GLM-5.2 runs at roughly a sixth of the cost of comparable Western tools. Where it is offered as open weights, it runs at whatever your own hardware costs. For a developer or a company weighing a monthly bill against a free download that performs almost as well, that is a genuinely hard question.

The strategy behind the giveaway

Handing away a frontier-class model is not charity; it is positioning. Open weights spread fast, become defaults inside companies and universities, and build an ecosystem that is very hard to dislodge. It is the same playbook that made earlier Chinese open models so influential, and it puts direct pressure on the closed, subscription-based approach of OpenAI and Anthropic.

There is a catch worth flagging. Using GLM-5.2 through Z.ai's hosted API means the calls are subject to Chinese data law, which is a non-starter for some enterprises handling sensitive code. But the whole point of open weights is that you do not have to use the hosted service, you can run the model entirely on your own infrastructure and sidestep that concern.

Pressure from below

The Western labs are fighting a two-front war. Above them, rivals push the frontier of raw capability. Below them, open-weights models like GLM-5.2 keep asking an awkward question: how much is a closed model worth when a free one does most of the job? ZCode is the latest, and one of the sharpest, versions of that question yet.