Product Signal

NVIDIA Launches Nemotron 3.5 Lightning

NVIDIA expanded the Nemotron 3 family with Nemotron 3.5 Lightning and introduced NeMo Switchyard for multi-model routing in agent workloads.

By All We Need Editorial System

NVIDIA has released Nemotron 3.5 Lightning alongside NeMo Switchyard, a system for routing agent tasks across models. The central change is not merely another open model: NVIDIA is bringing long-running agent efficiency and runtime dispatch into one product strategy.

A move beyond a single model

NVIDIA describes Nemotron 3.5 Lightning as a model for long-running agentic AI workloads and as the most efficient model in its class. The release extends the earlier Nemotron 3 line, which NVIDIA had positioned as an open, efficiency-focused hybrid Mamba-Transformer MoE family with up to one million tokens of context. The new release therefore pushes the family further toward workloads that require long context, repeated tool use and multi-step execution.

Efficiency increasingly depends on routing

NeMo Switchyard is designed to send agent requests to different models based on their capabilities, limitations and cost profiles rather than fixing every task to one model. NVIDIA’s technical explanation says those trade-offs vary by task. Earlier NVIDIA results for Nemotron 3 Super reported higher throughput than selected comparison models under specific B200, long-context and optimized-serving configurations. That suggests agent economics can depend on model selection and serving configuration as much as model size.

The competitive surface for open deployment widens

For enterprises, the value of open models is usually not limited to downloadable weights; it also includes control over cost, data location and task routing. By launching Lightning and Switchyard together, NVIDIA is attempting to address that deployment layer. However, NVIDIA says the existing nemo-relay-switchyard plugin is experimental, disabled in default builds and scheduled to be replaced by a native plugin. The direction is clear, but the routing layer is not yet fully mature.

What to watch next

Watch for three observable signals: whether NVIDIA expands availability through weights, APIs or major cloud platforms; whether it publishes cost, latency and task-success comparisons on real agent workloads; and whether the native Switchyard plugin reaches a stable release with third-party deployment evidence. Those outcomes would test whether the efficiency claim becomes reusable enterprise infrastructure.

Sources