NVIDIA launches Nemotron 3.5 Lightning, a 30B MoE model with 3B active parameters

Nemotron 3.5 Lightning is an open Mixture of Experts (MoE) model of 30 billion parameters that activates only 3 billion per forward pass. Its stated goal is to run agents that operate continuously and must complete many specialized tasks quickly. NVIDIA says the model provides up to 4× output speed compared with similarly sized models, a claim that refers to token generation rate, not answer quality.

In the PinchBench benchmark, which simulates agent workloads with 10,000 tasks, Lightning reached 86% accuracy and completed the full set 35% faster than Qwen3.6 35B at comparable accuracy. PinchBench is not an industry-standard benchmark such as MMLU or HumanEval; it specifically tests tool-calling and result verification scenarios, the type of work NVIDIA targets. The comparison to Qwen3.6 35B is relevant because both belong to the same deployment-size category.

The model is designed for post-training specialization through NVIDIA NeMo on domain-specific data, tools, workflows and policies. NVIDIA reports accuracy gains on cyber, code, legal and energy tasks, four verticals where agents perform many tool calls, validation and task delegation. The core use case is long-lived agents that spend most of their time executing rather than planning, and Lightning is aimed at that stage.

The small active size (3 billion) enables deployment on anything from an NVIDIA DGX Spark workstation (formerly Project DIGITS) to full data-center clusters. NVIDIA has not published exact VRAM requirements or latency figures for different configurations, making it difficult to assess deployment cost relative to alternatives such as Llama-3.1-8B or Qwen2.5-7B, which also run on modest hardware.

Alongside the model, NVIDIA releases NeMo Switchyard, an open-source library for routing between models. The concept is to use large, expensive frontier models for planning and complex reasoning, and offload massive, task-specific execution to Lightning. This “right model for the right task” approach is gaining traction in agent architectures, but Switchyard is still in an early version and its routing flexibility in practice—such as dynamic context-aware routing versus static rule-based routing—remains unclear.