Nemotron 3.5 Lightning & NeMo Switchyard Explained (2026)

On August 11, 2026, NVIDIA announced two new products addressing the cost and speed challenges of always-on AI agents: the open model "Nemotron 3.5 Lightning" and the open-source model routing library "NeMo Switchyard." Lightning is a lightweight model built to execute specialized tasks quickly and at scale, while Switchyard automatically routes each request to the optimal model. Together they form a stack designed to run always-on agents faster and cheaper without sacrificing accuracy.

When an AI agent runs continuously, routing every task — from code generation to code review, tool use, and security monitoring — through a top-tier frontier model keeps accuracy high but lets cost and latency balloon. Not every task demands frontier-level precision; simple checks and repetitive patterns rarely need an expensive model. NVIDIA's "system of models" concept addresses this inefficiency: a large model like Nemotron 3 Ultra plans and orchestrates the overall workflow, while a lightweight model like Nemotron 3.5 Lightning executes specialized tasks such as code review, tool use, and security monitoring quickly and at scale. This division of labor is designed to cut cost and time without giving up accuracy. The Nemotron 3 series underpinning this concept is covered in our Nemotron 3 hardware requirements guide.

Nemotron 3.5 Lightning is a Mixture-of-Experts (MoE) model with 30B total parameters and 3B active parameters, described as a "30B-A3B" configuration. It uses a hybrid architecture that interleaves Mamba-2 layers with MoE layers, combined with some Attention layers, and it was trained via distillation from the larger Nemotron 3 Ultra model as an extension of the Nemotron 3 series.

- Architecture: a hybrid design interleaving Mamba-2 and MoE layers, combined with some Attention layers
- Training: trained via distillation from the larger Nemotron 3 Ultra model
- Performance claims (NVIDIA): up to 4x higher output throughput and 30% shorter task completion time versus comparable open models, with "frontier-level accuracy" claimed on the agentic PinchBench benchmark
- License: OpenMDW-1.1, a permissive license that allows commercial use; training data and methodology are also released to the extent the license permits
- Availability: Hugging Face (BF16 as nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-BF16 and a quantized NVFP4 version), ModelScope, OpenRouter (with a free tier), build.nvidia.com, and NIM microservices
- Customization: can be post-trained on proprietary domain data using NeMo

NVIDIA's official materials mention serving on a single H100 or a DGX Spark, as well as running on RTX PCs, DGX Station, and Jetson. These are all sized for single-node deployment, and the model supports long-context use and structured tool calling. Because only 3B of the 30B total parameters are active in this MoE design, inference computation is lighter, which generally favors throughput. NVIDIA has not published specific VRAM requirements, so check official sources before planning a deployment.

Among 30B-class models, Meta's Muse Glimmer 30B, covered in our Muse Glimmer 30B hardware requirements guide, is a natural comparison. The two models are close in total parameter count but differ in architecture and licensing direction.

NeMo Switchyard is an open-source model routing library built for AI agents. It analyzes each request and automatically routes it to the optimal model across open, proprietary, and NVIDIA-built models. Priorities for quality, latency, and cost can be customized, and no rebuild of the existing application is required. According to NVIDIA's internal measurements, routing through Switchyard can "cut completion cost to about one-third compared to Opus 4.8 alone while maintaining frontier-level accuracy." It is available on GitHub, with integrations into partner platforms underway.

NVIDIA also published measured results from several partner companies.

Nemotron 3.5 Lightning is available on Hugging Face in BF16 and NVFP4 formats, as well as via ModelScope, OpenRouter (with a free tier), build.nvidia.com, and NIM microservices. NeMo Switchyard is an open-source library on GitHub, with integrations into partner platforms in progress. Full details are in NVIDIA's official blog post.

Routing every task in an AI agent workflow through a top-tier model is reassuring for accuracy, but costs add up quickly once the agent runs continuously. Matching model choice to task difficulty, and automating that choice through a routing layer, is a cost-optimization idea worth considering for SMBs and contract development teams running AI agents on a limited budget.

Nemotron 3.5 Lightning is a 30B-total/3B-active MoE model built to execute specialized tasks quickly and at scale, while NeMo Switchyard routes each request to the optimal model to compress cost. Together, the two products form a new stack for running always-on agents faster and cheaper without sacrificing accuracy, and the pattern of a large model planning while a lightweight model executes looks set to become a guiding idea in agent design.