Particle.news

NVIDIA Releases Nemotron 3.5 Lightning and NeMo Switchyard

The company positions a 30‑billion-parameter mixture-of-experts model with an open routing SDK to lower agent costs, reduce latency, speed multi-model deployments.

Overview

  • NVIDIA published Nemotron 3.5 Lightning and the NeMo Switchyard router on Tuesday, August 11, 2026, and made the model and code available on Hugging Face, ModelScope, OpenRouter, build.nvidia.com and GitHub.
  • Nemotron 3.5 Lightning is a 30 billion parameter mixture-of-experts model that activates about 3 billion parameters per token and uses multi-token and speculative decoding to boost throughput.
  • NeMo Switchyard is an open-source, provider-agnostic routing library (written in Rust) that evaluates each agent step and directs work to the cheapest or fastest model that meets quality, latency or cost targets.
  • NVIDIA and early partners report large cost and latency gains—vendor benchmarks claim up to 4x faster token generation, about 30% faster agentic task completion, and examples of 30–74% cost reductions—but those figures come from internal and partner tests and need independent validation.
  • Journalists report NVIDIA is developing Nemotron 4 with a largest model aimed at 1 trillion-plus parameters, but that project is still in training with no firm release date; enterprises can run Lightning locally and post-train it, yet production use will require per-step checks, escalation rules and independent benchmarks to guard accuracy and safety.