NVIDIA build.nvidia.com (NIM API)

The model pages officially state the free tier is rate-limited to 40 RPM + 10,000 requests/day. The older mechanism was 1000 API credits on signup (1 credit ≈ 1 request), unlockable up to a 5000-credit cap and 200 RPM with work-email verification — whether that credits pool still stacks on top of the rate limit is a matter of dispute as of 2026-09, so treat your logged-in account panel as the source of truth.

Verified live on 2026-09-01: both DeepSeek V4 variants have a Free Endpoint — deepseek-v4-pro(-0813) (1.6T total params/49B active, MoE + Hybrid Attention) and deepseek-v4-flash(-0731) (284B total params/13B active) — both with a 1M context and max output of 384K tokens, OpenAI-compatible (integrate.api.nvidia.com/v1), governed by the NVIDIA API Trial Terms, and sharing the same free tier as every other model in the catalog (the model page's rateLimits field literally says "Up to 40 rpm + 10,000 requests per day") — this isn't a V4-exclusive unlimited grant, and there's no evidence of a time-limited takedown (the page's DEPRECATION attribute turned out to be routine version rotation, not a shutdown). The credits mechanism itself has an unresolved dispute: a 2026-08-25 developer-forum thread still requests quota increases under the 1000→5000-credit framing with no official denial, but several third-party sites claim credits were dropped in favor of pure rate-limiting (those sites contradict themselves internally, with no official confirmation); the official model pages only show rate-limit copy, no mention of "credit" — check your logged-in account panel to be sure. No credit card is required to register, but activating an API Key usually needs SMS phone verification, and mainland China +86 numbers frequently fail to receive the code (multiple threads on NVIDIA's China-region forum and community sites) — try a different number or file a manual appeal.

⚠️The full-size DeepSeek-V4-Pro (1.6T, 1M context) Free Endpoint is the biggest highlight right now — well suited to a small number of high-value calls, like a one-off pass over a large repo or a long document. DeepSeek's own timeline: preview of V4 on 2026-04-24, Flash went GA on 07-31, Pro went GA on 08-13. No credit card required is accurate, but SMS phone verification is the main practical hurdle for users in mainland China.