Do NVIDIA's NIM cloud endpoints work free? Blocked on a missing key

Ledger entry: nvidia-api-catalog. User ask: "investigate how this thing is used... tell me if you managed to get it working for free and what type of tasks or usages you managed to do with it."

Hypothesis

A free NVIDIA Developer Program API key gives working, OpenAI-compatible chat access to hosted open models at no cost, good enough for prototyping tasks.

Hard constraints

  • Synthetic, non-sensitive prompts only. NVIDIA's Terms license user content to NVIDIA: no Betec, client or personal material.
  • Creating an NVIDIA account / key is the user's action. If NVIDIA_API_KEY is absent, do all key-free work (docs, catalog page, terms, quotas), then stop and report the exact blocker. Do not sign up, and do not use a key from anywhere else.
  • Never print or commit the key.

Cases (run in order, record each as pass / fail / blocked with evidence)

  1. Access: obtain model list via GET https://integrate.api.nvidia.com/v1/models. Pass = HTTP 200 and list saved to outputs/.
  2. Cost: after N calls, confirm no billing/credit prompt appeared; record quota/rate-limit headers or the documented limits. Pass = zero charge evidence plus limits written down.
  3. Compat: same prompt through the plain curl and through the OpenAI-compatible request shape the lab's openrouter-cli uses (base URL swap only). Pass = identical request code works with only URL/key changed.
  4. Task matrix, 2 open models of different size, 3 trials each, fixed prompts in source/prompts.md: (a) summarize a 300-word synthetic text, (b) extract JSON from a messy sentence, (c) write a 20-line Python function, (d) one Hebrew Q&A, (e) a tool-call/JSON-mode request. Record latency, tokens, validity (does JSON parse, does code run).
  5. Limits: 20 sequential calls in a minute; record first 429 and its message.
  6. Terms: quote the current terms lines on production use and data licensing, with URL and access date.

Control

The same task matrix on one OpenRouter free model, so "usable" is relative to something.

Deliverables

insights/RESULTS.md: did it work free (yes / no / blocked), which task types worked, which failed, limits, terms caveats. demo/index.html: the crux use case, "send a prompt to a hosted open model with a free key and see the answer", built from real recorded outputs only.