What Is NVIDIA Nemotron?
NVIDIA Nemotron is a family of downloadable AI models for developers building agents, rather than a single consumer chat application. Its catalog spans reasoning, visual understanding, speech, retrieval and safety workloads. Selecting Nemotron therefore starts with a particular checkpoint and task, not a family-wide promise that every model accepts the same inputs. Official family overview.
The family is most relevant when a team wants to customize model behavior or operate inference within its own infrastructure. Model weights, customization libraries and optimized serving software are separate pieces of that decision. NVIDIA NeMo supplies customization tools; NVIDIA NIM packages inference deployments. An external inference provider adds its own access and billing terms.
Choose a Model for the Actual Input
| Model lane | Task to consider | Important boundary |
|---|---|---|
| Nemotron 3 Nano | Text reasoning and smaller agent components | A text checkpoint does not acquire vision from the Nemotron family name |
| Nemotron 3 Super | Long-context text agents and tool use | Its existing MTP and long context predate Super VL |
| Nemotron 3 Nano Omni | Workflows combining visual material and speech | Its audio support should not be assumed for Super VL |
| Nemotron 3 Ultra | Larger text reasoning and complex agent orchestration | A separate capacity and deployment choice |
| Nemotron 3.5 Super VL | Agents working with documents, images and video alongside tools | The post-trained visual model produces text; it is not an image or video generator |
These are selected relevant checkpoints, not the entire catalog. The Nano, Super, Nano Omni and Ultra cards define their respective inputs and requirements.
For the October release, the Nemotron 3.5 Super VL version assessment covers the changes from the earlier Super model, BF16 versus NVFP4, and the deployment decision.
Access and Operating Costs
Downloading an openly available checkpoint does not include the GPU capacity or hosted service needed to run it. Keep three cost decisions separate:
- Self-hosted weights: provision memory for the chosen precision, runtime state, visual inputs and concurrency. The active parameter count measures sparse computation; it is not the complete weight footprint.
- Hosted inference: use the selected provider's current model identifier, rate, quota and data-handling terms. This page does not quote a universal Nemotron token price.
- Enterprise serving and support: evaluate NIM and NVIDIA AI Enterprise separately from the downloadable model. The October technical article describes NIM packaging and enterprise support as distinct parts of deployment. Serving overview.
Customization and License Boundaries
The official Super VL cookbook links serving notebooks, financial-document LoRA, multimodal reinforcement learning and a visual investigation agent. These resources are useful starting points for a team with a defined task and suitable infrastructure.
Openness varies by artifact. The 3.5 Super VL cards identify OpenMDW-1.1, while the earlier Super card names the NVIDIA Nemotron Open Model License. The Super 3.5 training documentation also lists unpublished training blends and unavailable dependencies. Published weights and recipes do not establish that the full reference training run can be reproduced from public components.
Avoid If
- The requirement is a ready-made chat subscription. Nemotron selection involves a model and an access route; this page does not describe a bundled consumer workspace.
- The target task requires native speech input but the selected checkpoint is Super VL. Nano Omni is a different model lane, and its audio capability is not transferable by name.
- The deployment plan budgets only for active parameters. Sparse activation reduces computation while the complete model still needs a suitable memory configuration.
- The project requires a fully reproducible 3.5 reference training pipeline using only released components. The training recipe documents material gaps.
Sources & Verification
- Last verified: October 18, 2026.
- Scope: model-family identity, selected model lanes, access routes and adoption constraints.
- Method: source-based research from the official materials linked above. No hands-on inference, fine-tuning or performance testing was performed for this update.
- Version-specific evidence: Nemotron 3.5 Super VL release assessment.
