ModelHubby

← All guides · provider record →

NVIDIA NIM (build.nvidia.com) — free tier setup

What you get free

Permanent free access to 100+ hosted models (including Llama 3.1 405B and DeepSeek-R1) behind one OpenAI-compatible endpoint, rate-limited by requests rather than tokens. Per the catalog: "40 requests/minute (cheahjs); 'Rate-limited (no daily token cap)' (mnfst); '~40 RPM' per model." The gotcha: this is the highest-friction signup of the free tier set — NVIDIA Developer Program membership plus phone verification — and there is no official public rate-limit page, so the 40 RPM figure is repo-sourced, not NVIDIA-published.

Friction check
Credit card no
Phone verification yes
Org/membership yes — NVIDIA Developer Program
Est. time to first key ~15m (5 steps)
Trains on your prompts unknown — a third-party source states prompts to one free-trial model are logged by NVIDIA; NVIDIA's own policy for build.nvidia.com not verified

Get a key

  1. Go to build.nvidia.com and sign in / create an NVIDIA account.
  2. Join the NVIDIA Developer Program and complete phone number verification when prompted (per repo sources).
  3. Open any hosted model's page on build.nvidia.com (e.g. meta/llama-3.1-405b-instruct).
  4. Use the API-key control on the model page to generate a key — keys are "issued from build.nvidia.com model pages" per the catalog (exact button label unconfirmed — verify on first signup).
  5. Copy the key (typically nvapi-... prefixed — unconfirmed — verify on first signup).

First request

Base URL confirmed from the official API reference: https://integrate.api.nvidia.com, POST /v1/chat/completions. Expected success shape: a JSON chat.completion object with choices[0].message.content.

export NVIDIA_API_KEY="YOUR_KEY"
curl https://integrate.api.nvidia.com/v1/chat/completions \
  -H "Authorization: Bearer $NVIDIA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta/llama-3.1-405b-instruct",
    "messages": [{"role": "user", "content": "ping"}],
    "max_tokens": 16
  }'

Limits that will bite you

  • ~40 RPM per model (repo-sourced; no official public rate-limit page found — expect drift).
  • No daily token cap reported, which makes it unusually good for sustained throughput — but treat that as unverified operating behavior, not a guarantee.
  • shared_pool: unknown — whether the 40 RPM is per model or account-wide is unconfirmed; probe before depending on parallel-model fan-out.
  • Hosted endpoints tend to be context-window limited vs self-hosted NIM; catalog caps: meta/llama-3.1-405b-instruct 128K context / 4,096 max output, qwen/qwen2.5-72b-instruct 8,192 max output.
  • Tool-calling support varies by model and was not confirmed.

Key rotation runbook

  • Second key: generate another key from a model page on build.nvidia.com (multiple concurrent keys unconfirmed — verify on first signup).
  • Revoke: key management/revocation UI location unconfirmed — verify on first signup.
  • ModelHubby integration: export NVIDIA_API_KEY_2="..." — the key pool picks it up automatically; state machine cools down rate-limited keys per-key.

ModelHubby usage

pnpm mh show nvidia-nim
pnpm mh probe --live --provider nvidia-nim --apply