GLM-5.2 draws praise for rivaling frontier models in open-source release

GLM-5.2 tops Terminal Bench 2.1 and achieves 80% pass rate on internal financial benchmarks, with users comparing it to Opus 4.8 and GPT-5.5. It runs 120 tok/s on two Blackwells and is free on HuggingFace Inference Providers, with NVIDIA releasing a NVFP4 quant.