NVIDIA lists GLM-5.2 on NIM with 753B parameters, a 1,000,000-token context window, and commercial or noncommercial access, giving developers a hosted path to test Z.ai's long-context model. NVIDIA's documentation says the third-party model targets long-horizon reasoning, coding, and agentic workflows, while Z.ai and Hugging Face describe it as an MIT-licensed open model. Z.ai's model card and blog describe GLM-5.2 as built around Dense-Sparse-Alternating layers plus IndexShare sparse attention, which the company says reduces per-token FLOPs by 2.9x at 1M context length. The immediate practitioner value is lower friction for experiments that would otherwise require hard-to-source multi-GPU infrastructure. The caution is that million-token context still needs workload-specific evaluation for latency, retrieval quality, safety, and cost before production use.