GLM-5.2 open-source model draws praise for reasoning and cost

GLM-5.2, a sub-trillion parameter open-source model, has been praised for reasoning close to Opus 4.8 and GPT-5.5, with a 80% pass rate on a financial benchmark and top spot on Terminal Bench 2.1. It runs at 120 tok/s on two Blackwell boxes and is available free via HuggingFace Inference Providers and on Perplexity's Agent API.