NVIDIA Build Inference Endpoints

Inference Endpoints on NVIDIA Build provide free inference access to more than 100 AI and machine learning models hosted on NVIDIA's DGX Cloud. Developers can use these endpoints to experiment with and integrate powerful models without upfront cost — no credit card required — via an OpenAI-Compatible API endpoint, as part of the broader NVIDIA Build Overview platform experience. [NVIDIA Build Overview] [NVIDIA Build Free API 100+ Models 2026]

The inference endpoints are organized into four skill domains, each grouping models by their primary use case: [NVIDIA Build Overview]

The following models are highlighted as available through the inference endpoints: [NVIDIA Build Overview]

• z-ai glm-5.2 — Tagged as *Agentic AI* (+3 additional tags)
• nvidia nemotron-3-ultra-550b-a55b — Tagged as *Agent* (+4 additional tags)
• nvidia cosmos3-nano-reasoner — Tagged as *Physical AI* (+7 additional tags)
• deepseek-ai deepseek-v4-pro — Tagged as *MoE* (+3 additional tags)

Inference Endpoints are one of several core capabilities offered within NVIDIA Build. They complement:

• Agentic Skills — Pre-built capabilities that agents can call, enabling models accessed via endpoints to be wired into agentic workflows.
• NemoClaw — NVIDIA's safe agent execution environment, which can leverage inference endpoints for secure, access-controlled AI operations.
• Blueprint Collection — Workflow templates and code samples that can be built on top of inference endpoint-backed models.
• DGX Station — Step-by-step playbooks (including NemoClaw setup) that may utilize inference endpoints as part of personal AI agent configuration.

• Inference is offered free for the listed models. [NVIDIA Build Overview]
• Models span a broad range of AI disciplines, from general machine learning to physical AI (robotics/simulation) and specialized developer tooling. [NVIDIA Build Overview]
• The endpoint catalog includes models from both NVIDIA and third-party providers (e.g., DeepSeek, z-ai). [NVIDIA Build Overview]

Every model in the catalog exposes the /v1/chat/completions endpoint in the exact format defined by the OpenAI API. Existing code written against openai.ChatCompletion.create() works by changing two variables: base_url becomes https://integrate.api.nvidia.com/v1 and the API key becomes the one generated on build.nvidia.com. [NVIDIA Build Free API 100+ Models 2026]

Model identifiers follow the convention provider/model-name visible on each model's catalog page (e.g. meta/llama-3.1-70b-instruct, deepseek-ai/deepseek-v3.1, zhipuai/glm-5.1). [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)

The catalog at build.nvidia.com/models contains more than 100 models organized by family, size, and specialization, covering general-purpose LLMs, coding, reasoning, multilingual, and vision tasks. [NVIDIA Build Free API 100+ Models 2026]

Specialized models for vision (LLaVA, Cosmos, NV-CLIP), retrieval (NV-EmbedQA, NV-Rerank), speech (Riva, Parakeet), biology (ESMFold), and physics simulation are also available. All hosted models are open-weight, published under permissive licenses (Apache 2.0, MIT, Llama Community License). [NVIDIA Build Free API 100+ Models 2026]

• Credits: Enrolling grants 1,000 initial credits; up to 4,000 additional credits can be requested via the developer forum (5,000 total). Credit consumption varies by model size. [NVIDIA Build Free API 100+ Models 2026]
• Rate limits: 40 requests/minute per model at the account level; exceeding this returns HTTP 429 with a Retry-After header. [NVIDIA Build Free API 100+ Models 2026]
• Context sizes: Not reduced from the originating lab's standard — Llama 3.1 70B keeps 128k tokens, Kimi K2 keeps 200k, DeepSeek V4 keeps 64k. [NVIDIA Build Free API 100+ Models 2026]
• After credits run out: API returns HTTP 402 on paid models; some smaller models remain accessible via base daily quotas. [NVIDIA Build Free API 100+ Models 2026]
• Data usage: NVIDIA states it does not use prompts or responses to train models; requests are processed in stateless mode. [NVIDIA Build Free API 100+ Models 2026]

Production note: The free tier is suited to prototyping and benchmarking, not production traffic. For production, NVIDIA recommends on-premise NVIDIA NIM container deployment (free for up to 16 GPUs for developer program members) or DGX Cloud with dedicated SLAs. [NVIDIA Build Free API 100+ Models 2026]

Because the endpoint is OpenAI-compatible, it integrates trivially with AI-first IDEs such as Cursor, Zed, OpenCode, and Hermes. The standard procedure is to enter https://integrate.api.nvidia.com/v1 as the base_url and the nvapi- key as the api_key in the IDE's custom AI provider settings. [NVIDIA Build Free API 100+ Models 2026]

For more details on NVIDIA NIM Framework Integrations and how these endpoints compare with paid alternatives, see Free vs. Paid AI Model Access: NVIDIA NIM vs. $20/month Subscriptions.

[NVIDIA Build Free API 100+ Models 2026]