NVIDIA NIM GPU Compatibility

NVIDIA NIM (Inference Microservices) packages AI models as portable containers that can be deployed on a range of NVIDIA GPU hardware. The NVIDIA Build catalog exposes a NIM Container GPUs filter that lets developers narrow the model list by the specific GPU each NIM container is validated to run on. [NVIDIA Build Models Catalog]

This page documents the GPU tiers surfaced in that filter, their significance for deployment planning, and how they relate to model availability across the catalog.

The following GPUs appear as discrete filter options in the NIM Container GPUs section of the NVIDIA Build catalog. [NVIDIA Build Models Catalog]

Note: Filter counts reflect the number of NIM containers explicitly tagged for each GPU at the time the catalog was observed. A single model may appear under multiple GPU tiers. [NVIDIA Build Models Catalog]

The diagram below shows how GPU tier selection fits into the overall NIM deployment flow.

• Hopper dominates: The H100 80GB HBM3 has the highest individual model count (16), reflecting its role as the primary data-center training and inference GPU in the current generation. [NVIDIA Build Models Catalog]
• Blackwell arrives: The B200 already carries 15 tagged models, signalling early Blackwell adoption in the NIM ecosystem. [NVIDIA Build Models Catalog]
• Ada Lovelace (L40S) holds parity with B200: The L40S — a PCIe-form-factor Ada card — matches B200 at 15 models, making it a practical on-premises and workstation-class option. [NVIDIA Build Models Catalog]
• H200 is close behind H100: With 14 models, the H200 (HBM3e) is nearly equivalent to the H100 in catalog coverage, suitable for memory-intensive large models. [NVIDIA Build Models Catalog]
• A100 SXM4 80GB remains relevant: Thirteen models are still tagged for the A100, preserving compatibility for existing Ampere infrastructure such as the DGX Station. [NVIDIA Build Models Catalog]

GPU compatibility interacts directly with other catalog dimensions:

• Free vs. downloadable: The catalog lists 77 Free Endpoint models and 108 Download Available models. [NVIDIA Build Models Catalog] Downloadable models are the ones where GPU tagging is most actionable, since they run on the operator's own hardware. See Free vs. Paid AI Model Access and the NVIDIA Build Inference Endpoints page for endpoint-type details.
• Model tiering: Choosing the right GPU tier is a core part of the Model Tiering Strategy — matching model size and memory footprint to available VRAM.
• Inference providers: When self-hosted GPU infrastructure is unavailable, developers can fall back to cloud NVIDIA NIM Inference Providers such as Deepinfra (34 models), OpenRouter (29), Together AI (23), GMI Cloud (15), and Lightning AI (7). [NVIDIA Build Models Catalog]
• Agentic workloads: Several recently added models — including GLM-5.2 and large MoE models — target agentic reasoning (Agentic Skills) and place high memory demands that favor H100/H200/B200 tiers. [NVIDIA Build Models Catalog]

The GPU-tagged models span publishers including NVIDIA (76 models), Meta (11), Google (6), Mistral AI (6), and Qwen (5). [NVIDIA Build Models Catalog] Publishers optimise their NIM containers for different GPU families; NVIDIA's own models cover the broadest range of hardware tiers.

For a broader view of what is available across the catalog, see NVIDIA NIM Free Models Guide and 80+ Free AI Models at NVIDIA Build.

While the catalog does not publish an explicit per-GPU use-case breakdown, the NIM Use-Case Taxonomy (NIM Use-Case Taxonomy) categories prominent in recently added GPU-tagged models include Drug Discovery (13), Image-to-Text (10), RAG (9), Speech-to-Text (9), and Code Generation (8). [NVIDIA Build Models Catalog] Memory-heavy use cases (multimodal, long-context MoE) naturally gravitate toward the H100/H200/B200 tier.