NVIDIA Build IDE Integration

NVIDIA Build exposes every model in its catalog through an OpenAI-Compatible API at the base URL https://integrate.api.nvidia.com/v1. Because modern AI-first IDEs already speak the OpenAI protocol, plugging them into NVIDIA NIM (Inference Microservices) requires changing only two values: the base_url and the api_key. The result is free access to more than 100 open-weight models — hosted on DGX Cloud — directly inside the editor, with no credit card required. [NVIDIA Build Free API 100+ Models 2026]

The OpenAI-Compatible API makes the integration trivially portable. The following editors are explicitly confirmed to work with the NVIDIA Build Inference Endpoints: [NVIDIA Build Free API 100+ Models 2026]

The IDE sends a standard POST /v1/chat/completions request. The NVIDIA NIM layer routes it to the requested model running on DGX Cloud and streams tokens back in the familiar OpenAI format. [NVIDIA Build Free API 100+ Models 2026]

The following procedure applies specifically to Cursor; the same fields appear under different menu paths in other editors. [NVIDIA Build Free API 100+ Models 2026]

1. 1.Open Settings → Models → Custom.
2. 2.Select OpenAI Compatible as the provider type.
3. 3.Fill in the fields:

1. 4.In the model list, manually add the identifiers you want, for example:

• meta/llama-3.1-70b-instruct
• deepseek-ai/deepseek-v3.1
• zhipuai/glm-5.1

1. 5.Save the profile. The models appear immediately in Cursor's model selector alongside Claude and GPT. [NVIDIA Build Free API 100+ Models 2026]

The same logic used by IDEs internally can be reproduced directly in any script. The OpenAI-Compatible API means the official openai Python package works without modification beyond two variables: [NVIDIA Build Free API 100+ Models 2026]

from openai import OpenAI

client = OpenAI(
    base_url="https://integrate.api.nvidia.com/v1",
    api_key="nvapi-XXXXXXXXXXXXXXXXXXXXXXXXXXXX"
)

response = client.chat.completions.create(
    model="deepseek-ai/deepseek-v3.1",
    messages=[
        {"role": "user", "content": "Explain in two lines what a MoE is."}
    ],
    temperature=0.6,
    max_tokens=512,
    stream=False,
)

print(response.choices[0].message.content)

Model identifiers follow the convention provider/model-name visible on each model card at build.nvidia.com/models. Switching models means changing only that one string — no other code changes are needed. [NVIDIA Build Free API 100+ Models 2026]

Not all catalog models are equally suited to coding and agentic IDE workflows. The following families support function calling in the standard OpenAI format and are well-suited to completion, refactoring, and chat agents: [NVIDIA Build Free API 100+ Models 2026]

• Llama 3.1 70B / 405B (Meta) — robust function calling, 128k context
• Nemotron-3-Super (NVIDIA) — strong on structured tool use; see NVIDIA NIM Framework Integrations
• GLM-4 / GLM 5.1 (Zhipu AI) — multilingual coding and tool use, agentic workflows
• Kimi K2 (Moonshot AI) — 200k context, ideal for large codebase analysis
• DeepSeek V3.1 / V4 — code generation and mathematical reasoning
• Mixtral 8x22B / Qwen 2.5 72B — function calling with strong structured output

Smaller models (under 7B parameters) and vision-only models either lack function calling or implement it only partially. [NVIDIA Build Free API 100+ Models 2026]

Understanding the free vs. paid model access boundaries prevents surprises during development: [NVIDIA Build Free API 100+ Models 2026]

• 40 req/min is rarely hit in interactive IDE use (chat, completions) but can become a bottleneck for automated evaluation scripts.
• Exceeding the rate limit returns HTTP 429 with a Retry-After header. [NVIDIA Build Free API 100+ Models 2026]
• Exhausting paid-model credits returns HTTP 402; some smaller models retain base daily quotas after credits are depleted. [NVIDIA Build Free API 100+ Models 2026]
• The free tier is not recommended for production traffic with real users; for that, NVIDIA points to on-premise NIM containers or DGX Cloud enterprise plans. [NVIDIA Build Free API 100+ Models 2026]

The integration is nearly invisible to existing OpenAI-targeting code. Minor differences to watch for: [NVIDIA Build Free API 100+ Models 2026]

• Model names use the provider/model-name convention rather than short aliases like gpt-4o.
• The logprobs advanced parameter may behave differently on some models.
• Streaming, function calling, and structured outputs work identically to the OpenAI specification for the supported models listed above.
• The JavaScript openai npm package (npm install openai) supports baseURL and apiKey overrides in exactly the same way as the Python client.

For a broader look at what the endpoint offers beyond IDE tooling, see NVIDIA Build Inference Endpoints and the NVIDIA NIM Free API Guide.

The Model Tiering Strategy consideration for IDE-connected free inference suggests four high-value scenarios: [NVIDIA Build Free API 100+ Models 2026]

1. 1.Model benchmarking — run the same prompt set across ten models in a single session to compare quality and latency before committing to one for a project.
2. 2.Agentic prototyping — build tool-calling agents with Llama 3.1, Nemotron, or GLM-4 without paying per iteration; relevant to Agentic Skills workflows.
3. 3.Long-document analysis — Kimi K2's 200k-token window handles entire codebases or books without chunking.
4. 4.Student and research projects — the catalog provides access to models that would otherwise require thousands of euros in local GPU hardware.