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

Pre-registered criteria: source/CRITERIA.md. Prompts: source/prompts.md. Demo: demo/index.html.

Verdict: "works free?" = BLOCKED (not yes, not no)

NVIDIA_API_KEY is not set on this machine (checked; ~/.config/api-keys/env has only GEMINI_API_KEY, and the lab .env only OPENROUTER_API_KEY). Creating the key (sign in at build.nvidia.com, "Get API Key", email + NVIDIA account) is the user's action, so I did not sign up and did not use another key. Single ask: sign in to build.nvidia.com in Chrome (the browser is NOT signed in), or create a free NVIDIA Developer key and put NVIDIA_API_KEY=... in ~/.config/api-keys/env. Then run, from outputs/: BASE_URL=https://integrate.api.nvidia.com/v1 API_KEY_VAR=NVIDIA_API_KEY MODEL=<id> LABEL=nvidia_<id> python3 run_matrix.py for two models (e.g. openai/gpt-oss-20b small, nvidia/nemotron-3-ultra-550b-a55b large; both in the listing), plus the 20-call burst. No NVIDIA chat answer was observed, so no success is claimed.

Per case

  1. Access: PASS (listing only). GET /v1/models returned HTTP 200 with 80 models and needed no key. Files: outputs/case1-models.json, case1-models.headers, case1-cmd.log. Caveat: the list mixes chat, embedding, vision and safety models; being listed does not mean callable (see surprise below). Not tested which are callable.
  2. Cost: BLOCKED. No authenticated calls, so no billing or rate-limit-header evidence. Unauthenticated chat gives HTTP 401 Header of type authorization was missing with no ratelimit headers (case2-nokey-chat.txt, .headers). Documented position: see case 6.
  3. Compat: BLOCKED (NVIDIA side). Shape is the standard OpenAI /v1/chat/completions + Bearer header; the 401/410 error replies show the endpoint parses it, but that is not the pass criterion. outputs/run_matrix.py is written to take only BASE_URL/key/model; it ran unchanged-shape against OpenRouter (control). I did not run the lab's openrouter-cli itself (interactive chat tool; it also only swaps OPENROUTER_ENDPOINT).
  4. Task matrix: NVIDIA BLOCKED; control done. Control, qwen/qwen3.8-27b:free on OpenRouter, 3 trials per task (outputs/matrix_control_or_qwen3.8-27b.json, control-run.log, control-summary.json):
    • a summary 3/3 (54-58 words), b JSON 3/3, c code 3/3 (extracted and executed, asserts passed), d Hebrew 3/3, e tool call 3/3. Median latency 0.9s (tool) to 24s (code); it is a reasoning model, so code and summary use ~1700-2100 completion tokens.
    • 4 of 15 trials needed retries after HTTP 429 "temporarily rate-limited upstream" (shared free pool); latencies exclude the 15s backoffs. So "free" OpenRouter was usable but flaky.
    • Only one model on the control side, one model-size; judged by my own simple checks (summary only length-checked, not for quality).
  5. Limits (20 calls/min): BLOCKED (needs key). Not run on the control either; it is not part of the plan for the control.
  6. Terms: DONE (accessed 2026-10-03). Files in outputs/.
    • Source: NVIDIA API Trial Terms of Service, https://assets.ngc.nvidia.com/products/api-catalog/legal/NVIDIA%20API%20Trial%20Terms%20of%20Service.pdf (terms-api-trial.txt): "NVIDIA will provide you with access to the API Service for limited trial purposes only and without use of the API Service or Generated Content in production." Also: "you may only use the API Service for internal testing ... not in production". Section 1.4 mentions "Credits", i.e. a credit-limited trial, but I did not find the credit amount or rpm figure in the fetched pages.
    • Data: you grant NVIDIA "a worldwide license with respect to User Content and Generated Content, to use, host, store, reproduce, modify, create derivative works ... during your API Service session solely to provide you with the API Service and as otherwise described in this Agreement"; 2.3 says NVIDIA won't store or use content after the session unless disclosed; NVIDIA may log content for security/fraud/abuse (3.3). Since the plan assumed a broader licence, this is milder than expected but still: synthetic data only.
    • Docs FAQ, https://docs.api.nvidia.com/nim/docs/product (terms_https___docs_api_nvidia_com_nim_docs_product.txt): Developer Program access to NIM API endpoints is "for prototyping"; production "requires an NVIDIA AI Enterprise license" (from $4500 per GPU per year); the free option is a 90-day AI Enterprise trial. The homepage https://build.nvidia.com/ says "Free inference with leading models".
    • Not retrieved: build.nvidia.com/terms_of_use is a client-rendered page (empty via curl), and the nvidia.com agreements page is a JS shell. Fallback was the PDF above; the PDF is an older-looking document and may not be the exact text a new key holder is shown.

Surprises

  • NVIDIA's own quickstart model meta/llama-3.1-8b-instruct now returns HTTP 410 "reached its end of life on 2026-08-26" (case1b-chat-nokey.txt), and it is also gone from the listing. The quickstart doc (updated "about 2 months ago") still names it, so copying the docs verbatim will fail. Catalog churn is a real cost for hard-coded model ids.
  • The listing works with no key, and nvidia/nemotron-3-ultra-550b-a55b, nvidia/nemotron-3.5-lightning and google/gemma-4-31b-it also appear as :free on OpenRouter, so a keyless-for-NVIDIA route to the same model families exists as a fallback.

Deviations (all recorded)

  • No-key state means cases 2, 3, 4 (NVIDIA side) and 5 were not run. Nothing substituted for them; the control stands alone.
  • Synthetic text in source/text_a.txt is 255 words, not 300.
  • Control run 1 (outputs/run1_max1500_noretry/) failed: max_tokens 1500 truncated the reasoning model's code answers with empty content, and 429s hit with no retry (4/15 passed by that log). I changed max_tokens to 6000 and added up to 5 retries with 15s backoff, then re-ran (run 2, the one reported). Both are kept.
  • Demo's NVIDIA section shows only recorded listing/401/410 results; its answer table is the OpenRouter control, labelled as such.

History: interruption and resume

  • The first run was interrupted by a usage limit and resumed. Earlier outputs were kept; nothing was redone.
  • After the resume I tried the authorised route: an opus subagent opened a new Chrome tab on build.nvidia.com. Result: Chrome is not signed in there (a "Login" button shows; the API keys page stays on a loading spinner). "Generate API Key" opens a dialog asking to agree to the NVIDIA API Trial Terms of Service; the agent closed it without agreeing, did not click Login, and stopped as instructed. No key was created, ~/.config/api-keys/env was not touched (no NVIDIA_API_KEY line). The BLOCKED verdict above stands. Notes also in insights/PROGRESS.md.
  • Surprise: the earlier assumption that Chrome was signed in to NVIDIA is false.