NVIDIA AI Releases Nemotron 3.5 Lightning: A 30B Open MoE with 3B Active Parameters, and NeMo Switchyard Model Router
| Source: MarkTechPost
Tags: NVIDIA, Nemotron 3.5, MoE, NeMo Switchyard, open-source, agentic AI, Mamba-2
NVIDIA releases Nemotron 3.5 Lightning — an open 30B mixture-of-experts model with only 3B active parameters, 1M-token context window, and 4x faster output than comparable models — deployable on a single H100 under the permissive OpenMDW-1.1 license, alongside NeMo Switchyard, an open-source agent routing library.
Details
NVIDIA released two interrelated tools for agentic AI infrastructure. Nemotron 3.5 Lightning is a 30B MoE model using a hybrid Mamba-2 plus attention architecture with only 3B active parameters per inference step and a 1M-token context window. NVIDIA reports 4x faster output speed than similarly-sized dense models and 30% faster completion of 10,000 PinchBench tasks versus Qwen3.6 35B at comparable accuracy. The model targets the execution layer of agentic systems — tool calls, result validation, subagent delegation — while larger Nemotron models handle orchestration and planning. Pre-trained on over 20 trillion tokens using an NVFP4 recipe, it is the smallest member of the Nemotron 3 family. NeMo Switchyard, released alongside it, is an open-source routing library that directs individual agent workflow steps to the most capable and cost-effective model available automatically, enabling hot-path optimization without application code changes. Licensing and deployment access are notable: Nemotron 3.5 Lightning ships under OpenMDW-1.1 with open weights, open training data, and training recipes. It runs on a single H100 or DGX Spark (GB10), putting it within reach of individual developers and startups. Early customizers include CrowdStrike, Harvey, CodeRabbit, and Lila Sciences across cybersecurity, legal, coding, and life sciences.