NVIDIA’s free model catalog isn’t a secret so much as something that’s grown large enough to notice.
Since opening its hosted NIM (NVIDIA Inference Microservices) endpoints to developers back in 2024, the company has kept expanding what’s available through build.nvidia.com/models.
The list has now crossed 80 production-grade models, including MiniMax M2.7, DeepSeek 3.2, Kimi 2.5, GLM 5.1, GPT-OSS-120B, and Sarvam-M, all reachable through a single OpenAI-compatible API key at no cost.
How the Free Tier Actually Works
Signing up runs through the NVIDIA Developer Program, requires only an email (and sometimes phone verification), and needs no credit card. Once inside, generating an API key under account settings produces a string prefixed nvapi- that unlocks the full catalog, shown once so it needs saving immediately.
New accounts start with roughly 1,000 inference credits, with up to 5,000 available on request, capped at 40 requests per minute.
Some models in the catalog don’t draw from the credit pool at all and are simply free outright, while others consume credits per call.
Because the API mirrors OpenAI’s request format, switching an existing project over is mostly a matter of swapping the base URL to NVIDIA’s endpoint and pointing it at a model ID like moonshotai/kimi-k2.6 or deepseek-ai/deepseek-v4 rather than rewriting application logic.
Notably, the hosted versions aren’t stripped-down variants. Kimi keeps its full 200k context window, Llama 3.1 70B retains its 128k context window, and NVIDIA says it doesn’t use submitted prompts or responses for training. Instead, it processes requests statelessly and keeps logging to a minimum.
The Catch, and Why NVIDIA is Doing This at All
The free tier is genuinely useful for prototyping and genuinely unsuited to anything with real users attached.
The rate limit and credit ceiling make it impractical for production traffic, and NVIDIA’s own guidance steers developers who outgrow it toward one of two paid paths: downloadable NIM containers self-hosted on up to 16 GPUs for free under the developer license, or metered DGX Cloud capacity with actual service-level guarantees once volume gets serious.
Latency in practice runs 600 to 1,500 milliseconds to first token depending on region, throughput sits in the 30 to 90 tokens-per-second range, and both degrade further during high-traffic periods like US peak hours or right after a popular new model lands in the catalog.
Developer reports testing the free tier against tools like Cursor and Cline have found some models slow enough to be impractical, requiring real trial and error to find one that performs consistently.
None of this is an oversight; the business model works exactly as intended. The free hosted tier and the self-hosted NIM containers run the same underlying software, so every developer who prototypes on build.nvidia.com naturally moves toward NVIDIA’s own GPUs and cloud infrastructure as projects scale. That makes the free API less of a gift and more of a deliberate funnel.
Used as intended, an evaluation layer to compare a handful of open models quickly before committing real budget anywhere, the free tier earns its spot in a developer’s toolkit. Used as a backend for a live product, the same rate limits that make it free are exactly what will break first.
Source: Medium (Coding Nexus), "NVIDIA Is Offering 80+ AI Models for Free via APIs"
