NVIDIA is offering free API access through build.nvidia.com, where its NIM platform, short for NVIDIA Inference Microservices, lets users work with more than 100 models from a single account. The lineup listed in the source includes Kimi, DeepSeek, MiniMax, Llama, Gemma, Qwen, and Mistral. The service runs on DGX Cloud, and its API format is compatible with OpenAI-style integrations.
Free quota starts at 1,000 credits
New accounts receive 1,000 inference credits by default. Users who need more can go to the API Keys page and click ��Request More” to apply for another 4,000 credits, bringing the total ceiling to 5,000 credits per account. The free tier also carries a 40 RPM rate limit, meaning up to 40 requests per minute, and requests sent too quickly may trigger 429 errors.
API keys can be configured with an expiration period of up to 12 months. The source also notes that the platform has been seeing heavy usage. In testing, timeout and 503 errors appeared at times, which the article attributes to server-side queueing rather than local setup issues.
Account setup requires four steps
The process begins with account creation on build.nvidia.com, followed by email verification. After logging in, users must complete phone verification through the on-screen prompt to unlock API access. According to the source, +886 mobile numbers are accepted, and the process does not require a VPN or a credit card.
The next step is creating an API key from the API Keys page. The article stresses that the full key string is shown only once. If it is not copied immediately, it cannot be retrieved later and must be regenerated. The final step is selecting a model, checking its Model ID, and inserting the generated API details into an application. Because the interface closely matches OpenAI’s format, the adjustment work is relatively limited.
Model list spans coding, reasoning and multimodal tasks
The source highlights several models to try first, including Kimi-K2.5, DeepSeek V3.2, MiniMax-M2.5 / M2.7, the Meta Llama family, and NVIDIA Nemotron. It also mentions Google Gemma, Alibaba Qwen, and Mistral. Based on the article, the available options cover long-context reasoning, code generation, math-focused workloads, and multimodal use cases.
Free use is limited to learning and prototyping
The FAQ in the source focuses on credits, rate limits, and access issues. If the initial credits run out, users can request more directly. If a key is lost after creation, it cannot be recovered. For repeated 429 or timeout errors, the article recommends checking request speed and adding exponential backoff to retry logic.
The source also states that the free tier is restricted to personal prototyping and learning use. It is not available for commercial or production deployment. Users seeking a higher rate limit than 40 RPM can apply through NVIDIA’s developer forum.