NVIDIA open-sources 30B MoE agent model Nemotron 3.5 Lightning, launches NeMo Switchyard router
On August 11 NVIDIA released a 30B MoE with 3B active parameters, outputting up to 4x faster than peers. NeMo Switchyard cuts multi-model workflow cost to roughly one-third of using Opus 4.8 alone.
On August 11, NVIDIA expanded the Nemotron 3 model family with Nemotron 3.5 Lightning, an open-weight 30B Mixture-of-Experts (MoE) model with only 3B active parameters, designed for high-volume, low-latency execution in always-on agents.
The model is trained and optimized for high-frequency tasks like tool calling, result validation, and sub-agent delegation, delivering up to 4x faster token generation and 30% faster task completion compared to peer open models in its class. Architecturally it incorporates speculative decoding, multi-token prediction, and DFlash and DSpark inference optimizations, with NVFP4 and BF16 quantization checkpoints for deployment across NVIDIA RTX PCs, DGX Spark, OEM GB10 systems, Jetson, DGX Station, GB300 workstations, and data centers.
NVIDIA also open-sourced NeMo Switchyard, an intelligent model-routing layer that dynamically routes agent workloads to the best-fit model based on cost, latency, and capability. Internal benchmarks show a Switchyard-routed multi-model system runs at roughly one-third the cost of Opus 4.8 alone while maintaining frontier-level task completion. Early adopters include Ramp, which cut task runtime 33% and cost 58%, and LangChain, which cut multi-turn agent costs 74% with negligible accuracy loss.
Nemotron 3.5 Lightning weights, training data, and recipes are released under OpenMDW-1.1. The model is live on Hugging Face, ModelScope, OpenRouter, and build.nvidia.com as an NVIDIA NIM microservice. Enterprise early testers include CrowdStrike, CodeRabbit, and Harvey; ecosystem partners include Siemens, Cadence, Cognition AI, and LangChain.