NVIDIA NIM is a genuinely free, no-credit-card cloud API — the easiest way to run AgentPlayground
without paying anything or installing local models. It's OpenAI-compatible under the hood, so it
reuses the same provider code as OpenAI, just pointed at a different base URL
(https://integrate.api.nvidia.com/v1).
- Go to build.nvidia.com and create a free account — no credit card.
- Generate an API key (starts with
nvapi-). - Rate limit is generous for personal use: roughly 40 requests/minute.
Enter it in the first-run setup wizard's NVIDIA card, or later in Settings → API Keys. If
NVIDIA is the only key you have, the app automatically makes it your default provider — you don't
need an Anthropic or OpenAI key to get a working install.
| Model | Label |
|---|---|
| meta/llama-3.1-8b-instruct | Llama 3.1 8B (free) — the default |
| qwen/qwen2.5-coder-32b-instruct | Qwen2.5 Coder 32B (free) |
| meta/llama-3.3-70b-instruct | Llama 3.3 70B (free) |
| deepseek-ai/deepseek-r1 | DeepSeek R1 (free) |
The default is deliberately the small 8B model — it's faster and burns far fewer free-tier
requests than the bigger models, which stay available in the picker for when you need more
capability. You can also type any other vendor/model id NVIDIA serves into the custom-model
field (e.g. mistralai/mixtral-8x7b-instruct-v0.1).
If you set an NVIDIA key, it's also used automatically as the free fallback for background work —
the quick-task router's model pick and the council/planner machinery — whenever no Anthropic key
is configured, so a free-tier-only install still gets those features working, not just chat.
NVIDIA reports an invalid key with a 403, not the 401 most providers use — the app's error
handling accounts for this specifically. See API key errors.
- Ollama (local, free) — the other free option, fully offline
- OpenAI — same underlying code path