FULL STORY

NVIDIA Launches Nemotron 3.5 Lightning and Routing Toolkit

NVIDIA released the Nemotron 3.5 Lightning model and NeMo Switchyard routing library. The model was subsequently deployed on major inference platforms for practical use.

2026-08-11 ~ 2026-08-12 · 2 episodes · 43 posts

Episode 1 · NVIDIA Unveils Nemotron 3.5 Lightning and NeMo Switchyard Router (2026-08-11, 41 posts)

On August 11, NVIDIA released the open-source Nemotron 3.5 Lightning model and the NeMo Switchyard routing library. Lightning is a Mixture-of-Experts (MoE) model with 30B total parameters and 3B active parameters, specifically designed for high-frequency execution in long-horizon agents. It supports a 1M context length, has open weights available on Hugging Face, and can be deployed on platforms like DGX Spark. NeMo Switchyard dynamically routes requests to the most suitable underlying model based on request signals, agent states, tool call results, as well as model capabilities, costs, and latency.

Confirmed

  • Model Architecture: 30B total parameters, 3B active (some sources indicate 31.6B/3.6B with a hybrid Mamba-Transformer architecture). It supports NVFP4 and BF16 precisions, features a 1M context length, and includes MTP, DFlash, and DSpark capabilities.
  • Open Source Content: Model weights and data are fully open and free to use on Hugging Face, capable of running on a single GPU.
  • Deployment Support: NVIDIA officially demonstrated deployment on DGX Spark, and the model is already live on the Baseten platform.
  • Routing Library: NeMo Switchyard is an open-source library supporting dynamic routing, accompanied by an official video explanation.

Unconfirmed

  • Specific performance benchmarks and comparative data (such as comparisons with gpt-oss-120b) were not provided with official sources in the materials.
  • Whether the model is an update from Nemotron 3.0 Nano was only mentioned by kuchaev and has not received official confirmation.

Why It Matters

This release marks NVIDIA's significant strategic push into Agentic AI. By leveraging an efficient MoE model and an intelligent routing library, it reduces the inference costs and latency of long-horizon agents, driving AI deployments from the edge to the cloud.

21 more related posts →

Episode 2 · Nemotron 3.5 Lightning Debuts on Inference Platforms (2026-08-11, 2 posts)

NVIDIA's Nemotron 3.5 Lightning is now available on Baseten and CoreWeave, offering 4x throughput and half cost for long-horizon agents.