GLM-5.2 open weights tested
Zhipu’s open MIT-licensed model, compared with GPT-5.5 and Fable 5 on several tasks — but not on Welsh.
The facts
| Developer | Zhipu AI / Z.ai |
|---|---|
| Launched | 13 June 2026 (GLM Coding Plan first, with open weights a few days later) |
| Architecture | MoE with around 750 billion total parameters (sources vary between 744B and 753B) / around 40 billion active; a one-million-token context window |
| Licence | MIT — a permissive open licence |
Welsh: no data from the developer
Material about GLM-5.2, including its marketing website, focuses on coding and agentic work, comparing it directly with GPT-5.5 and Claude Fable 5. No evaluation was published for Welsh, Celtic languages or lower-resource languages, and we found no specific claim about Welsh.
Nevertheless, the combination of open weights, an MIT licence and around 40 billion active parameters means it could, in principle, be further fine-tuned on Welsh data. That is the approach used by Jupiter-N with the Nemotron family. There is no evidence that anyone has tried this with GLM-5.2 yet; we will update the page if such a project appears.
CymraegBench v0.1 result
Overall score 68.42 · Welsh reasoning 68.62 · practical Welsh 68.22 · translation BLEU 57.07. GLM placed fourteenth, with unusually balanced performance between the two main categories. See the full leaderboard.