What is Nemotron 3.5 Lightning?

Nemotron 3.5 Lightning is Nvidia's open-weight model for autonomous agent workloads, released August 11, 2026. It's a hybrid architecture interleaving Mamba-2 and Mixture-of-Experts layers with attention — 30B total parameters with only 3B active per token — and Nvidia claims up to 4x faster output and 30% faster task completion than similar models. It runs on a single consumer GPU.

Is Nemotron 3.5 Lightning free for commercial use?

Yes. It ships under the OpenMDW-1.1 license, which allows commercial use, and Nvidia also released the training data and training recipes. Weights are available on Hugging Face, Nvidia's site, Ollama, LM Studio, and Cline.

What is NeMo Switchyard?

NeMo Switchyard is the open-source routing library Nvidia launched alongside Nemotron 3.5 Lightning on August 11, 2026. It directs AI requests to the most appropriate model — small local models for cheap steps, bigger models for hard steps — which is the deployment pattern Lightning is built for.

Quick Answer

What Is Nvidia Nemotron 3.5 Lightning? (August 2026)

Published: August 15, 2026

The Short Answer

Nemotron 3.5 Lightning is Nvidia’s free, open-weight model for autonomous agent workloads, released August 11, 2026. It’s a 30B-parameter hybrid MoE with only 3B active per token, interleaving Mamba-2 and MoE layers with attention. Nvidia claims up to 4× faster output and 30% faster task completion than comparable models — and it runs on a single consumer GPU.

Key Facts

	Nemotron 3.5 Lightning
Released	August 11, 2026
Parameters	30B total, ~3B active (hybrid MoE + Mamba-2 + attention)
License	OpenMDW-1.1 (commercial use allowed)
Openness	Weights + training data + training recipes
Hardware	Single consumer GPU
Claimed speed	Up to 4× faster output, 30% faster task completion
Where to get it	Hugging Face, Nvidia, Ollama, LM Studio, Cline
Companion release	NeMo Switchyard (open-source model router)

Why Nvidia Is Giving Away Models

Nvidia doesn’t sell tokens — it sells the hardware tokens run on. With Chinese labs (DeepSeek, Qwen, Z.ai, Moonshot) dominating open-weight leaderboards, Nvidia and Meta both moved in August 2026 to plant what CNBC called a “very firm flag” for US open-weight AI. A free agent model that runs great on RTX-class GPUs and DGX boxes is a hardware demand generator: every local agent loop is a GPU sale.

The Architecture Bet

Lightning’s Mamba-2 + MoE hybrid targets the specific economics of agents: long contexts read repeatedly, many small steps, latency compounding across a loop. State-space (Mamba) layers cut the cost of long-context token processing; the 3B-active MoE keeps per-step compute tiny. That’s how a 30B model claims 4× output speed on one consumer card.

Lightning + Switchyard: The Routing Pattern

Alongside the model, Nvidia launched NeMo Switchyard, an open-source router that sends each request to the most appropriate model. The intended deployment: Lightning handles the cheap, frequent agent steps locally, escalating hard steps to a frontier API. That mirrors how production teams already cut costs (see how to choose an AI model in 2026).

How It Compares

In the ~30B open agent-model class it lands directly against Meta’s Muse Glimmer 30B (Aug 10) and Qwen3.6-27B — comparison here: Nemotron 3.5 Lightning vs Muse Glimmer vs Qwen3.6-27B. Lightning’s distinctive angles: the most permissive openness of the trio (data + recipes, OpenMDW-1.1) and the speed-first architecture.

Who Should Care

Local-agent builders who want an always-on model on a 24GB GPU.
Cost engineers using small-model routing to cut frontier-API spend.
Enterprises needing license clarity — OpenMDW-1.1 plus published training data is unusually clean provenance.

Last verified: August 15, 2026.