GLM-5

Why we included this project

This repo is the open-weight model line behind Z.ai's coding and agentic products, so it suits teams building autonomous coding and software-engineering agents rather than people shopping for a wrapper or framework. The newest generation supports a solid 1M-token context that keeps long-running agent sessions coherent, and it posts strong scores on Terminal-Bench and SWE-bench Pro, landing within a few points of closed frontier models. Alongside the BF16 and FP8 weights, the repository includes the technical reports and deployment pointers for vLLM, SGLang, and Transformers, so you can self-host or call the hosted API. The IndexShare sparse-attention design, which cuts per-token FLOPs by 2.9x at 1M context, and the improved speculative-decoding layer are documented as well, which helps when you are estimating cost and latency at long context lengths. For repo generation, terminal automation, or other long-horizon agent work, having the weights, the benchmark numbers, and the architecture notes in one place saves you from rebuilding the design from the paper.