GLM-5.2 beats GPT-5.5 on coding, at one-sixth the cost
Zhipu's Z.ai just dropped the full MIT-licensed weights of GLM-5.2 on Hugging Face and ModelScope, and the numbers are the kind that make a US lab uncomfortable. It scores 62.1 on SWE-bench Pro, ahead of GPT-5.5 at 58.6, and sits at number two on Code Arena trailing only Claude Opus 4.8. It is a 744B mixture-of-experts model with 40B active per token, coding-first, with a 1-million-token context window. And you can download the whole thing.
The part that matters isn't that another model topped a benchmark. It's the combination. VentureBeat clocked it beating GPT-5.5 on multiple long-horizon coding benchmarks for roughly one-sixth the cost, and the Coding Plan starts at $10 a month. Open weights, MIT license, frontier coding, China-priced. That stack didn't exist six months ago.
There's one asterisk worth saying out loud: if you hit the hosted API instead of running the weights yourself, your data routes through China, which Techtimes flagged as a real consideration for enterprises. The escape hatch is the whole point of open weights though, you run it on your own metal and the data question goes away.
Here's my read. The open-weights frontier is now Chinese. Kimi, MiMo, and GLM have been climbing for months and this is the one where the closed labs can't wave it off as a cheap clone. The question for OpenAI and Anthropic stops being can we stay ahead on capability and becomes why should a developer pay six times more for weights they can't even hold. That's a much harder question to answer. Weights and details at https://huggingface.co/zai-org and https://z.ai.
← Back to all articles
The part that matters isn't that another model topped a benchmark. It's the combination. VentureBeat clocked it beating GPT-5.5 on multiple long-horizon coding benchmarks for roughly one-sixth the cost, and the Coding Plan starts at $10 a month. Open weights, MIT license, frontier coding, China-priced. That stack didn't exist six months ago.
There's one asterisk worth saying out loud: if you hit the hosted API instead of running the weights yourself, your data routes through China, which Techtimes flagged as a real consideration for enterprises. The escape hatch is the whole point of open weights though, you run it on your own metal and the data question goes away.
Here's my read. The open-weights frontier is now Chinese. Kimi, MiMo, and GLM have been climbing for months and this is the one where the closed labs can't wave it off as a cheap clone. The question for OpenAI and Anthropic stops being can we stay ahead on capability and becomes why should a developer pay six times more for weights they can't even hold. That's a much harder question to answer. Weights and details at https://huggingface.co/zai-org and https://z.ai.
Comments