GLM-5

Z.ai ships its flagship GLM line as genuinely open software: GLM-5 and its successors GLM-5.1 and GLM-5.2 are mixture-of-experts models in the 744B to 753B total parameter range with about 40B active per token, all under plain MIT with FP8 companion builds for efficient serving. The architecture adopts DeepSeek Sparse Attention to keep contexts around the 200K mark affordable, and the training recipe leans hard into software engineering and long-horizon agent work; Z.ai reports 77.8 percent on SWE-bench Verified for GLM-5. Adoption is unusually broad for a model this size, with quantized community builds of the newer checkpoints running in llama.cpp, Ollama, LM Studio, and Jan. Full-precision serving expects vLLM 0.19+, SGLang, or KTransformers with tensor parallelism across eight GPUs, while Z.ai's hosted API covers everyone else; the MIT terms leave commercial use, distillation, and redistribution unrestricted.