NVIDIA's Nemotron 3.5 Lightning checkpoint, released August 11, 2026, is a candidate for teams seeking efficient self-hosted agent inference. Its 30B total and 3B active parameters describe the architecture; they do not eliminate the need to account for all stored weights, runtime memory, and the hardware's supported compute path.
What the model card establishes
The official card describes a hybrid architecture combining Mamba, mixture-of-experts layers, and attention. It provides a hardware matrix, deployment recipes, and reported evaluations. The NVFP4 label refers to the released representation, while the runtime may execute a different path depending on the GPU.
That distinction is useful when comparing systems. A checkpoint can be compatible with several devices without delivering the same acceleration on all of them. Procurement decisions should use a validated recipe for the intended GPU generation, not just a claim that the model “supports NVIDIA hardware.”
Why active parameters do not equal memory requirements
For mixture-of-experts models, only a subset of parameters participates in an individual step, but the service must still make the required experts available. Context state, batching, and any speculative-decoding companion introduce additional capacity requirements. A small active count is therefore a reason to investigate efficiency, not a complete sizing specification.
Test at the intended concurrency. An internal assistant used by one developer and an agent pool serving many simultaneous tasks can have very different capacity needs even with identical weights. Track time to first useful output and time to complete the task, then repeat with representative long inputs. Reserve memory headroom instead of sizing directly to a model download.
Benchmark results belong to a configuration
NVIDIA supplies scores and evaluation recipes for the official checkpoint. Those are reported results with specific settings, rather than Nerova measurements or a promise about another harness. A tool parser, reasoning configuration, or quantization change can alter behavior in ways a top-level model name does not reveal.
For agent use, include schema correctness and recovery from tool errors in the evaluation. A cheap response that triggers retries or requires a human to repair the output can cost more per finished task than a slower alternative. Keep the environment and task set constant when measuring that tradeoff.
Review OpenMDW before distribution
The model is governed by OpenMDW 1.1. It should not be relabeled Apache 2.0 merely because the weights are downloadable. Preserve the governing license and review the intended use, modification, and distribution against its terms.
Nemotron 3.5 Lightning is most interesting when a team can validate a supported runtime and has enough demand to justify operating inference. Its efficiency case should be established with actual task completion and infrastructure cost, not inferred from the 3B active figure alone.