Zhipu ships GLM 5.2 with a 1M-token context and no benchmarks

Zhipu released GLM 5.2 on 13 June, rolling it out across every tier of its GLM Coding Plan with a usable one-million-token context window aimed at repo-scale agentic refactors.

It runs on the same 744-billion-parameter mixture-of-experts architecture as GLM 5 and caps output at 131,072 tokens. Unusually for a flagship launch, it ships with no benchmark numbers at all: no SWE-bench Verified, no LiveCodeBench, no HumanEval. Zhipu says the standalone API, the Z.ai chatbot, and MIT-licensed open weights all arrive next week.

Shipping a coding-first model to paying users before publishing a single score is a confident bet that hands-on results, not leaderboards, will sell it. The MIT open-weights drop next week is what the rest of the field will actually be watching.