GLM-5.1

Z.ai released GLM-5.1 as a mixture-of-experts model of 754 billion parameters, 40 billion of them active, under an MIT license. The Batch n°350, published by DeepLearning.AI on 24 April 2026, records what it did with that architecture: it took the lead on SWE-Bench Pro at 58.4%, against 54 to 57% for competing models, and ranked first on CyberGym at 68.7. It could hold a single task autonomously for up to eight hours through a plan-execution-evaluation loop.

The same coverage marks the limit. On GPQA Diamond, GLM-5.1 scored 86.2% while Gemini 3.1 reached 94.3%. Its results clustered in agentic and code work, and Z.ai raised its API prices by roughly 40% in the same period.

The lead did not last. Z.ai shipped GLM-5.2 on 16 June 2026, keeping the same shape (753 billion parameters, 40 billion active, MIT license) and adding a 1M token context. Artificial Analysis then placed GLM-5.2 first among open weights models and third overall on GDPval-AA, at 1524 Elo, ahead of the next open model MiniMax-M3 at 1408. Less than two months separate the two releases.

What GLM-5.1 fixes in the record is a moment: an open weights model under a permissive license holding the top score on an agentic coding benchmark while trailing on reasoning tests. The model that displaced it came from Z.ai as well, which leaves open how much of that convergence is a broader movement rather than one publisher's cadence.