GLM-5 lands in NVIDIA API Catalog

The interesting part here is not that GLM-5 exists; it is that NVIDIA is turning heavyweight open models into commodity infrastructure for agent builders. If the API experience holds up, this becomes another credible backend for coding workflows instead of just another benchmark darling.

• –NVIDIA's model card describes GLM-5 as a 744B MoE model with about 40B active parameters, roughly 205K context, tool calling, and structured JSON output.
• –The benchmark table is legitimately strong for an open model: 77.8 on SWE-bench Verified, 56.2 on Terminal-Bench 2.0, and 62.0 on BrowseComp.
• –The Kilo CLI / Kilo Code angle matters because agent builders care about provider optionality; swapping models is more valuable than being locked to one vendor's chat UI.
• –The caveat is that hosted availability does not guarantee great real-world agent behavior, so latency, rate limits, and tool-call reliability still decide whether teams keep it in rotation.