The Catalog is a self-service library of ready-to-deploy blueprints. Each card stands for a packaged resource, a GPU virtual machine, a Kubernetes cluster, a database, or a model service, that you can request from the console. The platform then provisions it into your project and lets you manage its lifecycle. This guide groups every blueprint by job and explains, in plain terms, what each one is and when you would reach for it.
Names of third-party models and tools (NVIDIA, Meta Llama, Qwen, Jupyter, Triton, Open-WebUI, Whisper) are vendor products and are kept as-is. Platform-specific and lab-specific identifiers have been replaced with neutral placeholders.
Compute and Environment Blueprints
These give you the raw place to run AI work: clusters, workstations, notebooks, and a database. Start here when you need infrastructure rather than a packaged model.
| Blueprint | What it is | Why you would use it |
|---|---|---|
| AI Kubernetes Cluster | A managed Kubernetes cluster with GPU-capable worker nodes, tuned for AI and ML workloads. | When you want to run containerized AI apps, model servers, or training and data pipelines on Kubernetes with GPU scheduling handled for you. |
| AI Workstation | A GPU-enabled virtual machine, with the option to pre-install common AI and ML frameworks. | For single-user development, experiments, and small fine-tuning runs where you want a ready GPU box without building it by hand. |
| Jupyter DLVM | A deep-learning virtual machine with PyTorch and JupyterLab installed under Conda. | For notebook-driven data science: exploring data, prototyping models, and sharing reproducible analysis. |
| Triton Inference Server | A GPU virtual machine running the NVIDIA Triton Inference Server. | When you need to serve models from several frameworks with batching and throughput optimizations on one server. |
| Managed PostgreSQL (Data Services) | A blueprint to deploy and manage PostgreSQL clusters as a service. | As the backing store for applications, metadata, and vector search with the pgvector extension. |
Model Serving Blueprints
These stand up a language or vision-language model as a running endpoint you can call. Pick one based on size, speed, and whether you need image input.
| Blueprint | What it is | Why you would use it |
|---|---|---|
| Deploy Model Endpoint | A guided blueprint that takes the guesswork out of serving a large language model. You pick the model and it provisions the rest. | The fastest path to a working inference API. Use it whenever you want a self-hosted chat or completion endpoint without writing deployment configs. |
| NIM-LLM Llama 3.2 1B Instruct | A small Meta Llama text model served through an NVIDIA NIM microservice. | For lightweight, low-latency chat and generation where a large model would be wasteful, such as classification, routing, or simple replies. |
| NIM-LLM Qwen 3.5 VL 3B Instruct | A compact Qwen vision-language model served through NIM. | For tasks that mix images and text at a small footprint, like describing a picture or answering questions about a screenshot. |
| Llama 3.2 11B Vision Instruct NIM | Meta Llama 3.2 Vision, a multimodal model for image reasoning, packaged as a NIM. | When you need stronger image understanding: reading charts, captioning, and document-image question answering. |
Retrieval Models for RAG
Retrieval-augmented generation needs two model types: one to turn text into vectors, and one to reorder search results so the best context reaches the model. These blueprints provide both.
| Blueprint | What it is | Why you would use it |
|---|---|---|
| llama-nemotron-embed-1b-v2 | An NVIDIA text embedding model. | To convert documents and queries into vectors so you can run semantic search and feed a RAG pipeline. |
| NVIDIA Embedding V5 | The NeMo Retriever text embedding model, built for high-quality retrieval. | A strong default for production RAG when retrieval quality matters and you want well-separated vectors. |
| NVIDIA llama-nemotron-rerank-1b-v2 | A reranking model that scores how likely a passage is to answer a given query. | To reorder retrieved chunks after the first search so the model sees the most relevant context first, which lifts answer accuracy. |
Perception Models: Text from Images and Audio
Much of your source content is locked inside scanned files and recordings. These models pull readable text out so it can be indexed and searched.
| Blueprint | What it is | Why you would use it |
|---|---|---|
| NVIDIA Nemotron-OCR | An optical character recognition model. | To extract text from scanned PDFs and images before indexing them into a knowledge base. |
| NVIDIA Parakeet CTC 0.6B ASR | A compact automatic speech recognition model, audio to text. | To transcribe calls, meetings, and voice notes so the words become searchable content. |
| Whisper Audio | A speech-to-text model for transcribing audio. | A second option for transcription, useful for multilingual audio and longer recordings. |
Access Layer: Gateway and User Interface
Once models are running, these blueprints put a clean front on them: one API for every provider, and a chat window for end users.
| Blueprint | What it is | Why you would use it |
|---|---|---|
| LiteLLM Gateway | A proxy that exposes many model providers (NVIDIA NIM, Bedrock, Hugging Face, Vertex AI, OpenAI, and others) behind a single OpenAI-format API. | So your application code talks to one API and one key store. You can switch or mix providers without rewriting the app. |
| Open-WebUI AI Interface | A deployment of Open-WebUI, a chat front end, backed by a pgvector database and pointed at the LLM Gateway. | To give people a familiar chat window over your private models, with built-in storage for history and documents. |
How the Pieces Fit Together
A full private AI stack usually pulls from several of these blueprints at once. You serve a text model with Deploy Model Endpoint or a NIM-LLM blueprint, add an embedding and a rerank model for retrieval, and store the vectors in Managed PostgreSQL with pgvector. OCR and ASR models widen what you can index by reading images and audio. The LiteLLM Gateway puts one API in front of all of it, and Open-WebUI gives users a place to chat. Heavier development and training happen on the AI Kubernetes Cluster, the AI Workstation, or a Jupyter notebook.
To deploy any of these, open the Catalog, find the card, and click Request. The blueprint collects a few inputs, then provisions the resource into your project where you can manage it from Instances.
Reference guide to the Private AI catalog blueprints. Vendor product names belong to their respective owners.
About The Author
DrJha