Zhipu GLM-5

GLM-5 is Zhipu AI's generational flagship model, released February 11-12, 2026 — a month after the company's Hong Kong IPO. It's a 744-billion-parameter mixture-of-experts model with ~44B parameters active per token, a 200K context window, and fully open weights under the MIT license. Reuters reported it approaching Anthropic's Claude Opus 4.5 on certain coding benchmarks and beating Google's Gemini 3 Pro on others, with 77.8% on SWE-bench coding and 92.7% on AIME math reasoning. Notably, it shipped with deep inference adaptation for Chinese domestic chips (Moore Threads, Cambricon) — part of Beijing's self-reliance push.

The honest picture: at 744B total parameters this is a datacenter-only self-host, and the MIT license covers weights but not training data or code. Zhipu's own numbers dominate the coverage — independent head-to-head verification against Claude Opus 4.5-class models is still thin. The GLM-5.2 successor (June 2026, 1M context) has already taken the spotlight, so GLM-5 is now the previous generation. But as a free, MIT-licensed near-frontier coding model, it remains one of the strongest open-weight options available.