A thing not enough people know about NVIDIA's hosted inference: the model catalog is a public endpoint. No account, no key:
curl -s https://integrate.api.nvidia.com/v1/models
Today that returns 102 models, and the DeepSeek section is the interesting part. The entries are deepseek-ai/deepseek-v4-flash-0731 and deepseek-ai/deepseek-coder-6.7b-instruct. DeepSeek-R1 and V3.1 — the ones a year of tutorials told you to point at — are not in the list anymore. The rotation tracked DeepSeek's own retirement of that generation, but nothing errors until you actually call the old ID.
The practical habit: hit that endpoint before writing any config that hardcodes a NIM model ID, and again whenever a call starts 404ing. It is faster than the web catalog, it is the ground truth for what your key can reach, and because it needs no auth you can even wire it into CI — a one-line check that every model ID in your repo still exists in the catalog will catch this whole class of rot before your users do.
Model lineups at every hosted provider churned this year; the catalog endpoint is the only place the change shows up the day it happens, not the month someone updates the docs.
Full picture of the NIM setup and its credit system: https://toolfreebie.com/nvidia-nim-free-api/
