About NIM Stats

NIM Stats is a public, no-login operational dashboard for the free NVIDIA NIM API endpoints. A collector sends a real chat-completion request to every tracked model on a fixed cadence, records what actually happened — time to first token, end-to-end latency, tokens per second, whether the call succeeded or timed out — and publishes the result. Nothing on this site is a vendor-published status claim; every number is a measurement taken from an ordinary API call, from outside NVIDIA's network, exactly as a developer's own client would experience it.

The question the site exists to answer is narrow and practical: of the free NIM endpoints available right now, which one will actually serve my request, and what will it cost me in latency and throughput if I switch to it? Trial-and-error against a dozen model IDs is slow and burns quota. A continuously probed fleet view answers it in seconds.

Each probe is a live chat-completion request against the endpoint, not a ping or a TCP health check. From the response stream the collector derives time to first token (TTFT), total latency, sustained throughput in tokens per second, and a congestion estimate. Successive samples are aggregated into per-model uptime, reliability, p95 and p99 latency, timeout rate, and a rolling incident history.

Endpoints are re-discovered daily, so models that NVIDIA adds appear automatically and models that disappear are retired. Retired endpoints are periodically re-verified rather than dropped permanently, so a transient error cannot keep a working model off the dashboard. Historical samples are pruned on a retention window; the site is a live operational view, not a long-term archive.

NIM Stats is not affiliated with, endorsed by, or operated by NVIDIA Corporation. "NVIDIA" and "NIM" are trademarks of their respective owner and are used here only to identify the service being measured. The site sells nothing, requires no account, and takes no payment.

Measurements are taken from a single vantage point on a fixed cadence, so they describe what that client observed at that moment. Your own latency will differ with geography, network path, prompt size, and how many other people are hitting the same free endpoint. Treat the fleet view as a strong prior for which endpoint to try first, not as a service-level guarantee.

The site is built and operated by Aathil Felix as an independent project. It is a personal engineering project rather than a registered company, and it has no commercial relationship with any model provider. Source-level questions, corrections, and requests to track an additional endpoint are welcome — see the contact page at https://nimstats.aathil.com/contact.