GPU Recommendations for Default NIM Models Innovation Release

Overview

From Hybrid Manager, there are two primary consumers of AI models:

  • PG.AI Knowledge Base (AIDB Postgres extension) for creating and maintaining AI Knowledge Bases.
  • Langflow for building AI pipelines and agentic flows.

Default NIM Models

Model typeNIM modelNVIDIA NIM documented resource requirements
Text completionnvidia/llama-3.3-nemotron-super-49b-v14 × L40S
Text embeddingsnvidia/llama-3.2-nemoretriever-300m-embed-v11 × L40S
Image embeddingsnvclip1 × L40S
OCRpaddleocr1 × L40S
Text rerankingllama-3.2-nv-rerankqa-1b-v21 × L40S

Minimum GPU Requirement

Based on the default models above, the minimum to run them concurrently is 8 × L40S GPUs.

Cloud Mappings

  • AWS EKS: recommend a node group with 2 × g6e.12xlarge nodes.
  • GCP GKE: recommend a node pool with 2 × a2-highgpu-4g nodes.

Note: GCP does not offer L40S GPUs. The recommended A2 nodes with A100 GPUs are supported and documented for the NIM models listed above.

Additional Supported NIM Models

Beyond the defaults above, AIDB and Langflow support the following additional NIM model variants, tested and confirmed to work with Agent Factory. Models marked Default on HM are the same ones already covered by the GPU sizing above; all others can be self-hosted with KServe and registered the same way, but aren't included in the minimum GPU requirement or cloud mappings on this page — consult the model's page on build.nvidia.com for its own resource requirements before deploying it.

Text Completion (nim_completions)

NIM modelDefault on HMDescription
meta/llama-3.1-8b-instructGeneral-purpose instruction-following chat model; solid default choice for most completion tasks.
nvidia/llama-3.3-nemotron-super-49b-v1Large reasoning/chat model for complex tasks needing higher-quality output.
nvidia/llama-3.3-nemotron-super-49b-v1.5Refined revision of llama-3.3-nemotron-super-49b-v1 with improved response quality.
nvidia/nemotron-3-nanoCompact, fast completion model — good for latency-sensitive workloads.
nvidia/nvidia-nemotron-nano-9b-v2Mid-size reasoning-capable model, balancing speed and capability.
openai/gpt-oss-120bLarge open-weight model for high-quality completions and agentic tasks.
openai/gpt-oss-20bSmaller open-weight model — faster, lighter-weight completions.

Text Embeddings (nim_embeddings)

NIM modelDefault on HMDescription
nvidia/llama-3.2-nemoretriever-300m-embed-v1General-purpose text embeddings for retrieval and RAG.
nvidia/llama-3.2-nv-embedqa-1b-v2Larger embedding model tuned for question-answering retrieval.
nvidia/nv-embedqa-e5-v5Embedding model tuned specifically for QA-style retrieval tasks.

Image Embeddings (nim_clip)

NIM modelDefault on HMDescription
nvidia/nvclipJoint text/image embeddings for multimodal search.

OCR (nim_paddle_ocr)

NIM modelDefault on HMDescription
baidu/paddleocrGeneral-purpose OCR text extraction from images.
nvidia/nemoretriever-ocr-v1NVIDIA's OCR model — an alternative to PaddleOCR.

Text Reranking (nim_reranking)

NIM modelDefault on HMDescription
nvidia/llama-3.2-nv-rerankqa-1b-v2Reranks retrieved passages by relevance to a query — use after an initial vector search to improve top-k precision.