Nvidia ships an open 'Lightning' model and a router for AI agent stacks

Nvidia shipped two pieces of an enterprise agent stack today: a 30-billion-parameter open model designed for narrow tasks, and a routing library that decides which model handles which request without forcing developers to rewrite their agent tools.

Nemotron 3.5 Lightning is a mixture-of-experts (MoE) model, an architecture that activates only a slice of its total parameters per token. Lightning has 30 billion total parameters but runs about 3 billion at inference, which keeps per-request compute low while keeping the model's range. Nvidia is selling it for the execution layer of long-running agent systems, the narrow-task slot where a smaller, faster model reads tool output, formats results, or runs follow-up calls after a larger reasoning model has framed the problem.

AI development is moving from "one giant model handles the whole job" toward tiered systems where a larger model plans or orchestrates and smaller specialized models execute. Nvidia is positioning itself for the second slot, the worker tier, and pairing the model with a router.

NeMo Switchyard is an open-source library that sits in front of an organization's mix of models, including open weights, proprietary APIs, and Nvidia's own. For each request, it picks the most capable or suitable model and hands the work over, without the agent's surrounding tool code having to know which model answered. The point is compositional flexibility: an enterprise can keep its existing Claude or GPT endpoints for planning tasks and route the narrow, repetitive calls to a Lightning-class model it controls.

Lightning is fully post-trainable through Nvidia NeMo, Nvidia's training framework, on a customer's own data, tools, and workflows. A legal-tech team can fine-tune the model on its own contract corpus and routing rules rather than prompt-engineering a generic frontier model. The model runs across RTX PCs, DGX Spark, DGX Station, and Jetson devices, plus workstations, edge, data center, and cloud. Nvidia places the family between the earlier Nemotron 3 Nano and larger reasoning models like Nemotron 3 Ultra, and frames the broader lineup as a continuously updated "software library" of open models.

The performance case is Nvidia's, not third-party benchmarked. Nvidia reports Lightning runs up to 4x faster than models in its class on output speed, and that this translates to roughly 30% faster completion of agentic tasks. On PinchBench, a long-horizon agentic benchmark, Nvidia claims frontier-level accuracy against rivals in the same category.

Customer adoption is also Nvidia-sourced. The company names CrowdStrike in cybersecurity, Harvey and Trajectory in legal, CodeRabbit and Baseten in code review, Lila Sciences in physical and life sciences, and Fastino Labs across software development, finance, and healthcare as early users. Useful as deployment-signal context, less so as independent validation. The open weights are already mirrored on consumer runtimes: LM Studio and pi.dev's OpenRouter listing make the model downloadable for local inference, which is consistent with the open-weight positioning but not a deployment win on its own.

Nvidia's developer blog also names Lightning alongside OpenClaw and Hermes Agent as harnesses supported by the NemoClaw open-source security and management stack, positioning the release as part of a wider integration surface for third-party agent frameworks rather than a model Nvidia ships against.

The watch item is whether the customisation story holds. If teams post-train Lightning on their own data and route a meaningful share of agent traffic through it, the launch reads as a serious bid for the worker tier of an emerging agent architecture. If deployments stall at logo-parking and the model ends up used as a generic open-weight alternative to Anthropic or OpenAI, the architectural framing is a marketing coat on a routine release. The first concrete test: a Nemo-post-trained Lightning model running production agent traffic at one of the named customers, and a third-party model integration in Switchyard beyond the demo stage.