NVIDIA will run GLM‑5.3 and Kimi‑K3 for you, for free. I read the fine print. Here are the three catches.

Description and use

NVIDIA’s free API offers 80+ models but is capped at 40 requests per minute, disallows production use, may use prompts for training, and frequently removes models without notice.

Problem

Users assume NVIDIA’s free AI model endpoint offers unrestricted, production‑ready access, but it imposes hidden limits, data‑usage clauses, and model volatility.

Outcome

The free NVIDIA endpoint is suitable for single‑threaded chat or low‑throughput agent use, but not for production or high‑fan‑out workloads due to a hard 40 RPM limit, data‑usage clauses, and frequent model removal.

Workflow

• 1. Register for NVIDIA Developer Program (free).
• 2. Obtain API key (prefix nvapi-).
• 3. Query https://integrate.api.nvidia.com/v1/models to get current model IDs.
• 4. Use the OpenAI‑compatible chat completion format to call desired model.
• 5. Implement exponential backoff for 429 responses; limit to 40 requests per minute per key.
• 6. Store prompts locally; avoid sending sensitive data that may be used for model training.
• 7. Monitor build.nvidia.com for model availability and deprecation notices.
• 8. If higher throughput or production use is needed, purchase NVIDIA NIM deployment.
• 9. Regularly check the Trial Terms of Service for updates on data usage and production clauses.

Evidence

"not in production" twice, a clause that lets NVIDIA use your prompts to improve its models

Source

Catch 1: 40 requests per minute, and no way up

Source

The API is ground truth, and it answers without a key: curl https://integrate.api.nvidia.com/v1/models | jq '.data | length' # 80 (15 Oct 2026)

Source