Use Case & Reason:
I am using the Hermes open-source tool for personal development. This tool requires making multiple sequential API calls to analyze e-commerce product listings (using the deepseekai-deepseek-v4-pro model). Even with a single monitoring task, the current 40 RPM limit is frequently exceeded, causing HTTP 429 errors that halt the workflow.
I have already reduced task frequency to the minimum and disabled concurrency. Raising the limit would make development and testing stable while staying within controlled, non-production usage.
I have already signed up and verified my account with my phone number.
This request is strictly for personal, non-commercial, and development purposes. Thank you for your support.
This is a NIM question rather than a TensorRT one, and rate-limit changes aren’t granted via forum posts in any category. There’s a pinned post in NVIDIA NIM > Access/Accounts that lays out the policy and the realistic options:
Worth reading end-to-end before any next step. I’m moving this thread over to that subcategory so the NIM team sees it in their queue:
Practical workaround in the meantime: since you’re already serializing the Hermes calls and disabling concurrency, adding an exponential-backoff retry on 429 plus a token-bucket rate limiter set just under 40 RPM (around 0.6 requests per second) around your API client will get you stable runs at the free tier without further code restructuring.
⚠️ Heads up: please rotate the API key whose last 4 characters you posted, from https://org.ngc.nvidia.com/setup/personal-keys. Even partial fragments are best kept out of public forum threads.