NVIDIA Build — Complete Overview

At the heart of NVIDIA Build is NVIDIA NIM (NVIDIA Inference Microservices), a hosted inference service running on NVIDIA's own DGX Cloud that makes over 100 AI models available through a single, unified endpoint — many of them at no cost. [80+ Free AI Models at NVIDIA Build] The catalog spans a wide range of providers and architectures: models like MiniMax, Kimi, DeepSeek, GLM, and GPT-OSS sit alongside NVIDIA's own models such as nemotron-3-ultra-550b-a55b and cosmos3-nano-reasoner, as well as third-party offerings like GLM-5.2 from Z.ai and deepseek-v4-pro from DeepSeek AI. [NVIDIA Build Overview] The full NVIDIA Build Models Catalog indexes 141 models spanning free endpoints, downloadable NIM containers, and NVIDIA NIM Inference Providers; the NIM Use-Case Taxonomy maps these across use-case categories. The catalog also includes LLMs (Llama 3, Mistral, Phi-3, Qwen), multimodal, code-focused, and specialized models for embeddings and reranking. [NVIDIA NIM Free Models Guide] GPU hardware requirements for running NIM containers locally are covered in NVIDIA NIM GPU Compatibility. The NVIDIA Open Model Ecosystem spans eight model families including voice, robotics, autonomous vehicles, and biomedical AI, all available free under commercial-friendly licenses. The NVIDIA AI Models 2026 Guide provides benchmarks, competitive comparisons, and hardware requirements across every NVIDIA model family.

A notable free-tier model is GLM-4 (by Zhipu AI), a bilingual Chinese/English model strong at reasoning, instruction-following, and coding tasks. Free accounts also receive an initial credit balance to test paid models in the catalog. [NVIDIA NIM Free Models Guide] The NVIDIA Build Inference Endpoints page documents the full set of free endpoints available across ML, accelerated computing, physical AI, and developer tools categories.

The practical significance of the catalog size becomes clear when contrasted with the dominant alternative: as Lo Zarantonello points out in 80+ Free AI Models at NVIDIA Build, most users currently pay around $20/month for access to a single AI model. NVIDIA NIM inverts that proposition entirely — a free account unlocks over 100 models simultaneously, making the Free vs. Paid AI Model Access: NVIDIA NIM vs. $20/month Subscriptions comparison decisively one-sided for developers who simply need capable, hosted inference. [80+ Free AI Models at NVIDIA Build] Developers can further reduce API spend by applying a Model Tiering Strategy — routing simpler tasks to free models and reserving premium models for high-stakes reasoning — yielding reductions of 40–70% on agentic workloads.

The platform's framing, as described by Lo Zarantonello, is explicitly transitional: on-device AI is positioned as the long-term destination, but until local hardware makes that fully viable, free cloud-hosted inference via a unified API is the practical bridge. [80+ Free AI Models at NVIDIA Build]

The engineering decision that makes NVIDIA NIM immediately accessible to the broadest possible developer audience is its OpenAI-Compatible API Endpoint. Every model in the catalog exposes the same request and response format as OpenAI's API, meaning any tool, library, or application already configured to use OpenAI — whether that is Claude Code, LangChain, CrewAI, AutoGen, or a custom Python script using the OpenAI SDK — can switch to NVIDIA's infrastructure by changing exactly one configuration value: the base URL. [80+ Free AI Models at NVIDIA Build] [NVIDIA NIM Free Models Guide]

The base URL for NVIDIA NIM is https://integrate.api.nvidia.com/v1, and credentials take the form of an API key (nvapi-xxxxxxxxxxxxxxxx) generated once from a free account at build.nvidia.com. The key is shown only once at generation time. Each model page in the catalog shows its exact model ID string, supported endpoints, pricing tier, and a live playground for testing before integrating. [NVIDIA NIM Free Models Guide]

This compatibility layer eliminates the primary source of switching friction. Developers do not need to learn a new SDK, rewrite API call logic, or restructure response parsing. The OpenAI-Compatible API Endpoint effectively makes NVIDIA NIM a drop-in replacement for any OpenAI-based workflow, while simultaneously offering access to a far broader and more cost-effective model catalog. [80+ Free AI Models at NVIDIA Build]

Beyond raw inference, NVIDIA Build provides a library of Agentic Skills — pre-built capabilities that an AI agent can call directly, organized into four domains. [NVIDIA Build Overview] Rather than implementing every tool or function from scratch, developers can compose agents from these ready-made building blocks, keeping each capability modular, testable, and independently maintained.

The AI and Machine Learning domain is by far the largest, with 143 skills, covering the core workloads most developers reach for first. Physical AI (37 skills) reflects NVIDIA's investment in robotics and embodied intelligence. The NVIDIA Physical AI Stack documents NVIDIA's integrated portfolio for physical-world AI spanning humanoid robotics, autonomous vehicles, and industrial AI — including GR00T N1.7 (a vision language action model for humanoid robots), Alpamayo R1 (a reasoning VLA model for Level 4 autonomous vehicles), and Cosmos 2.5 (world foundation models for synthetic training data generation). Accelerated Computing (25 skills) exposes GPU-accelerated processing capabilities, and Developer Tools (10 skills) rounds out the set with utility-level capabilities. [NVIDIA Build Overview]

Agentic Skills are designed to work in conjunction with NemoClaw, the platform's secure agent execution layer: NemoClaw governs which agents can invoke which skills, and under what conditions, enforcing access control at the level of individual capabilities. [NVIDIA Build Overview] This modular, policy-governed design reflects a broader philosophy on the platform — agents should be composable from well-defined, controllable primitives rather than monolithic, opaque systems.

NemoClaw is NVIDIA Build's answer to a fundamental challenge in production AI: how do you let an agent act autonomously while still maintaining meaningful control over what it can do and what it can see? The framework is organized around three pillars: safe agent execution (guardrails that limit unintended or unsafe actions), access control (policies governing which resources and data an agent can interact with), and data protection (shielding sensitive information throughout the agent's operation). [NVIDIA Build Overview]

On the NVIDIA Build portal, NemoClaw is the primary call-to-action — the "Build a Claw" prompt directs developers into a self-service onboarding flow. It is also the featured use case in the DGX Station getting-started playbooks, where setting up NemoClaw as a secure personal AI agent is explicitly the recommended first step for new users arriving with DGX Station hardware. [NVIDIA Build Overview]

The positioning of NemoClaw as a first-class offering on NVIDIA Build signals that NVIDIA is targeting not just hobbyist or prototyping use cases but production and enterprise deployments where security and auditability are non-negotiable requirements. Its integration with Agentic Skills, the Blueprint Collection, and DGX Station makes it the connective tissue that holds the platform's agentic story together. [NVIDIA Build Overview]

The Blueprint Collection is a curated library of workflows and code samples on NVIDIA Build that give developers structured starting points for building complete AI applications. [NVIDIA Build Overview] Where Agentic Skills provide individual building blocks and NVIDIA NIM provides inference, blueprints provide the higher-level architectural patterns — reference implementations that encode best practices and end-to-end structure so that developers can focus on customization rather than foundational scaffolding.

Blueprints draw on all four Agentic Skills domains and can leverage any model in the NVIDIA NIM catalog. They sit alongside NemoClaw and DGX Station playbooks in the platform's getting-started experience, and developers are directed to "Explore the Blueprint Collection" from the NVIDIA Build homepage. [NVIDIA Build Overview] The NemoClaw setup flow in particular — including the DGX Station playbook — is positioned as a recommended entry point that leads naturally into the broader blueprint library.

DGX Station is NVIDIA's personal AI computing platform — dedicated on-premises hardware for developers and researchers who want to run serious workloads locally rather than purely in the cloud. On NVIDIA Build, the DGX Station experience is surfaced through step-by-step playbooks accessible via a "Start Now" prompt. [NVIDIA Build Overview]

The playbooks tie DGX Station tightly into the rest of the NVIDIA Build ecosystem. Setting up NemoClaw as a secure personal AI agent is the highlighted first-use case, and from there developers are guided toward Agentic Skills, the Blueprint Collection, and the inference endpoints that can be leveraged as part of DGX Station-based workflows. [NVIDIA Build Overview] This integration means that DGX Station is not positioned as a standalone product but as the local execution layer for everything NVIDIA Build offers — a bridge between the cloud-hosted convenience of NVIDIA NIM and the latency, privacy, and cost benefits of on-premises compute.

Much of the accessible documentation about NVIDIA Build and NVIDIA NIM in this notebook comes from Lo Zarantonello, a software engineer, instructor, and writer affiliated with the GoPenAI publication on Medium. His article 80+ Free AI Models at NVIDIA Build, published on June 13, 2026, is a practical guide that frames the free model catalog as a direct challenger to the $20/month single-model subscription model. [80+ Free AI Models at NVIDIA Build] The NVIDIA NIM Free Models Guide, published by the MindStudio Team (May 3, 2026), complements this with a step-by-step walkthrough of account creation, API key generation, model selection, and integration into frameworks like LangChain, CrewAI, and AutoGen. [NVIDIA NIM Free Models Guide]

Together, these guides make the case that NVIDIA Build is not merely a NVIDIA marketing surface but a genuinely useful developer resource — one that lowers the barrier to AI development through free access, standard APIs, and a rich ecosystem of pre-built capabilities.

The architecture of NVIDIA Build is best understood as a stack, where each layer enables the one above it:

NVIDIA NIM provides the inference substrate — free, hosted, and universally accessible via the OpenAI-Compatible API Endpoint. Agentic Skills build on that substrate to give agents ready-made tools. The Blueprint Collection provides the architectural patterns that compose skills and inference into complete applications. Voice capabilities are available through PersonaPlex 7B, NVIDIA's open-source full-duplex speech-to-speech model with sub-second latency. Agentic LLM workloads are served by Nemotron 3, NVIDIA's flagship open-source LLM family built on a Hybrid Mamba-Transformer MoE architecture. Hardware-level token generation at scale is addressed by the Rubin Hardware Platform, NVIDIA's six-chip AI platform delivering inference at one-tenth the cost of Blackwell. NemoClaw wraps agent execution in a security and access-control layer. And DGX Station brings all of this to on-premises hardware for developers who need local execution. [NVIDIA Build Overview] [80+ Free AI Models at NVIDIA Build] [NVIDIA NIM Free Models Guide]

The result is a platform that addresses the full developer journey — from a first API call against a free model, through composing agentic workflows with pre-built skills, to deploying secure, production-ready agents on dedicated hardware — without requiring the developer to leave the NVIDIA Build ecosystem at any point.