GLM-5.2 review: 753B parameters, open-weights, Chinese-developed
Release date: mid-June 2026 (reported) | Status: Active | Licence: Open
A Chinese lab just put one of the biggest open-weights models yet on the public internet, and you can download the whole thing for free. That is the short version of GLM-5.2, released by Zhipu AI (now Z.ai) around the middle of June 2026.
Here is why an Australian business team should care. Most of the AI you can actually self-host comes with a trade-off: the powerful models are closed and rented by the token, and the ones you can run yourself tend to lag behind. GLM-5.2 muddies that line. It ships under an open licence, the weights are on Hugging Face, and at 753 billion parameters it sits at the top end of what anyone has released openly.
The catch is that the headline figures floating around for this model have been messy, and the numbers in the earlier draft of this review did not hold up against what the labs and trackers actually published. I have flagged those below rather than repeat them as fact. Treat the benchmark talk in this piece as directional, and check the source links before you bet a project on a specific score.
Benchmarks at a glance
A note before the table: several of these figures could not be confirmed against primary sources, and at least one (the context window) is plainly wrong in the original draft. I have kept the disputed numbers visible so you can see what was claimed, but read the "Context" column and the section below for the corrections.
| Metric | Score (as originally claimed) | Context |
|---|---|---|
| SWE-bench Pro | 51.4% (disputed) | Reported elsewhere as ~62.1%, see note below |
| MMLU | 85.2% (disputed) | Reported elsewhere closer to ~91.7% |
| Context window | 256K tokens (incorrect) | Actual headline figure is 1M tokens |
| Price (input) | $0.80 / 1M tokens (disputed) | Reported API price ~$1.40 / 1M |
| Price (output) | $2.40 / 1M tokens (disputed) | Reported API price ~$4.40 / 1M |
| Licence | Open | Self-hostable, MIT |
The licence is the one row I would stake money on. Z.ai released GLM-5.2 under an MIT licence with no regional restrictions, and the weights, including an FP8 variant, are downloadable from Hugging Face. That part is solid.
The 753B parameter question
The 753 billion parameter figure checks out. The official model page lists 753B params, and the major trackers agree.
What matters more than the raw count is how the model uses it. GLM-5.2 is a Mixture-of-Experts (MoE) design, so it does not fire all 753 billion parameters on every token. Only a slice is active at a time, which is what keeps inference affordable on hardware that does not cost a fortune. The earlier draft put the active count at roughly 60 billion per token and likened it to Llama 4; Artificial Analysis actually lists 40 billion active, in line with the previous GLM-5 and 5.1 releases. So the MoE point stands, but the 60B figure looks off, 40B is the number to use.
The practical upshot is the same either way: you get the knowledge capacity of a very large model without paying the full inference bill of one.
Chinese language performance
This is the part of the original review I would treat as a reasonable hunch rather than a measured result. A model built in China by a Chinese lab will almost certainly be strong on Mandarin tasks, classical Chinese translation, and China-specific knowledge, and that is consistent with how earlier GLM models behaved. But I could not find a benchmark that confirms the specific claim that GLM-5.2 beats Western models on Chinese reading comprehension. If your work involves Chinese-speaking customers, it is worth a look, just run your own evaluation before you commit, because the comparative edge here is reported, not proven.
Coding assessment
Coding is where the original numbers fall apart most, so read this section with caution.
The draft put GLM-5.2 at 51.4% on SWE-bench Pro and ranked it above Qwen 3 and Llama 4 but behind MiniMax M3 and Kimi K2.7-Code. That ranking rests on a figure that does not match the public record. Multiple outlets, including TechTimes, report GLM-5.2 scoring around 62.1 on SWE-bench Pro, ahead of GPT-5.5 and its own predecessor GLM-5.1, not behind a pack of rivals. The competitor scores quoted in the draft (Qwen 3, Llama 4, MiniMax M3, Kimi K2.7-Code) could not be verified and appear to be constructed, so I would not rely on that league table at all.
The claim that GLM-5.2 handles Python and Java well but struggles with JavaScript frameworks and Rust is also unconfirmed. No source breaks the model down by language at that level of detail, and the broader reporting actually points the other way: GLM-5.2 is being described as one of the strongest open-source coding models available right now, which sits awkwardly with a "struggles with" framing. Test it on your own stack before you write off any language.
For developers who want to dig in, Z.ai's code lives on GitHub (repo not independently confirmed at time of writing).
Verdict
Strip out the dodgy numbers and there is still a real story here: a Chinese lab has shipped a genuinely large, openly licensed, self-hostable model, and the early coding reports are strong. That is the maturing open-weights ecosystem doing what closed vendors keep saying can't be done cheaply.
If you need a capable open model you can run on your own infrastructure, or you specifically want a non-American option, GLM-5.2 belongs on your shortlist. Just verify the benchmarks that matter to your use case against the primary trackers rather than any single review, including this one, before you build on it.
Score: 7.9 / 10 (the author's original rating; note it was assigned against benchmark figures that did not hold up on review)