NVIDIA Build Model Catalog

NVIDIA NIM (Inference Microservices) originally launched as an enterprise product distributed as on-premises containers. In July 2024, NVIDIA shifted strategy and opened a hosted model catalog on DGX Cloud, free to any enrolled developer. The commercial rationale is twofold: encourage prototyping on NVIDIA endpoints before purchasing GPUs for production, and position NVIDIA as a neutral distribution hub for open-weight models — a counterweight to the closed ecosystems of OpenAI and Anthropic. [NVIDIA Build Free API 100+ Models 2026]

For a broader look at the inference endpoint infrastructure, see NVIDIA Build Inference Endpoints and the NVIDIA Build Overview.

As of April 2026, the catalog contains more than 100 models organized by family, size, and specialization. The table below summarizes the major language model families. [NVIDIA Build Free API 100+ Models 2026]

Beyond LLMs, the catalog also includes specialized models for: [NVIDIA Build Free API 100+ Models 2026]

• Vision — LLaVA, Cosmos, NV-CLIP
• Retrieval — NV-EmbedQA, NV-Rerank
• Speech — Riva, Parakeet
• Biology — ESMFold
• Physics simulation — additional domain-specific models

All hosted models are open-weight, published under permissive licenses (Apache 2.0, MIT, or the Llama Community License). NVIDIA does not silently modify the weights published by the originating research labs. [NVIDIA Build Free API 100+ Models 2026]

For more on GLM-4 and its place in the catalog, see the dedicated entity page.

1. Register for the NVIDIA Developer Program

Visit build.nvidia.com and click Get API Key. Create an account with email and password. No government ID or credit card is needed. Email verification completes in seconds. [NVIDIA Build Free API 100+ Models 2026]

2. Generate a Personal API Key

From build.nvidia.com/settings/api-keys, click Generate Key. The key appears only once (prefixed with nvapi-). Store it in a password manager or as the NVIDIA_API_KEY environment variable immediately. [NVIDIA Build Free API 100+ Models 2026]

3. Configure the OpenAI Client

The entire integration requires changing two variables from any existing OpenAI client code: [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 provider/model-name convention shown on each catalog card (e.g. meta/llama-3.1-70b-instruct, moonshotai/kimi-k2, zhipuai/glm-5.1). Switching models requires changing only that one string. [NVIDIA Build Free API 100+ Models 2026]

Key constraints to be aware of: [NVIDIA Build Free API 100+ Models 2026]

• Credit depletion: When free credits run out, the API returns HTTP 402 for flagship models. Smaller models may remain accessible via base daily quotas. Additional credits can be requested via the NVIDIA developer forum.
• Rate limiting: Exceeding 40 req/min returns HTTP 429 with a Retry-After header.
• Load variability: During US peak hours or new model launches, throughput can slow noticeably. NVIDIA recommends on-premise NIM containers for production workloads.
• No credit card required at any point to use the free tier.

Context window sizes are not reduced relative to the originating lab's published specifications. [NVIDIA Build Free API 100+ Models 2026]

For a detailed comparison of free vs. paid tiers, see Free vs. Paid AI Model Access: NVIDIA NIM vs. $20/month Subscriptions and the NVIDIA NIM Free Models Guide.

Most modern models in the catalog support function calling in the standard OpenAI format. Confirmed models include: [NVIDIA Build Free API 100+ Models 2026]

• Llama 3.1 70B / 405B
• Nemotron-3-Super
• GLM 4.7 and GLM 5.1 (see also GLM-4)
• Kimi K2
• Mixtral 8x22B
• Qwen 2.5 72B

Models under 7B parameters and some vision-only models either lack function calling or implement it partially. Each model card on build.nvidia.com lists supported features explicitly. [NVIDIA Build Free API 100+ Models 2026]

This capability is central to building Agentic Skills and multi-step AI pipelines without incurring per-prompt costs during prototyping.

Because the catalog exposes a standard OpenAI-Compatible API, integration with AI-first IDEs is straightforward: [NVIDIA Build Free API 100+ Models 2026]

Cursor example: Set name to NVIDIA NIM, base_url to https://integrate.api.nvidia.com/v1, api_key to your nvapi- key, then manually add model identifiers (e.g. meta/llama-3.1-70b-instruct). The models then appear in Cursor's model selector alongside Claude and GPT. [NVIDIA Build Free API 100+ Models 2026]

The right choice depends on the use case: Groq wins on latency; NVIDIA Build wins on catalog breadth; Together AI wins on post-depletion predictability; Hugging Face covers the long tail of experimental models. [NVIDIA Build Free API 100+ Models 2026]

See also the NVIDIA NIM Framework Integrations page for how NIM fits into broader orchestration stacks.

The 40 req/min rate limit and 5,000-credit ceiling make the free tier unsuitable for production applications with real user traffic, but well-suited for: [NVIDIA Build Free API 100+ Models 2026]

1. 1.Model benchmarking — Run the same prompt set across ten models to compare quality, latency, and projected cost before committing to one.
2. 2.Agent prototyping — Build and iterate on tool-calling agents (Llama 3.1, Nemotron, GLM 4.7) without paying per iteration.
3. 3.Long-document analysis — Kimi K2's 200k-token context handles entire books without chunking; 5,000 credits cover dozens of analyses.
4. 4.Academic / university projects — Access models that would otherwise require thousands of euros in GPU hardware to run locally.

For production workloads, NVIDIA recommends either on-premise NIM containers (free for up to 16 GPUs for developer program members) or purchasing dedicated DGX Station capacity with SLAs. [NVIDIA Build Free API 100+ Models 2026]

The endpoint is globally reachable. DGX Cloud serves primarily from US data centers, adding approximately 80–150 ms of network latency for European users. There is no geographic block or residency requirement. For chat and asynchronous generation workloads, this latency is negligible. [NVIDIA Build Free API 100+ Models 2026]

According to the official NIM API Catalog terms, NVIDIA does not use prompts or responses to train models. Requests are processed in stateless mode with no content logging beyond minimal operational data. For workloads involving sensitive data (healthcare, finance, PII), masking personal information before sending is still recommended regardless of the provider. [NVIDIA Build Free API 100+ Models 2026]

• NVIDIA Build — the portal and developer program
• NVIDIA NIM (Inference Microservices) — the underlying microservice technology
• NVIDIA Build Inference Endpoints — endpoint configuration details
• OpenAI-Compatible API — the API standard all catalog models use
• Free vs. Paid AI Model Access: NVIDIA NIM vs. $20/month Subscriptions — cost comparison
• NVIDIA NIM Free Models Guide — free model selection guide
• NVIDIA NIM Free API Guide — API key and quota guide
• GLM-4 — Zhipu AI's model family in the catalog
• Agentic Skills — using catalog models in agentic pipelines
• DGX Station — on-premise hardware for production NIM deployments
• Model Tiering Strategy — choosing the right model size for a task