Nvidia ships Nemotron 3.5 Lightning, a 30B/3B-active open MoE for agents

Nvidia has released Nemotron 3.5 Lightning, a 30B-parameter MoE that activates 3B parameters per token, under the OpenMDW-1.1 licence, alongside NeMo Switchyard, an open-source router that directs agent requests to different models.

Key points

• Model ID: NVIDIA-Nemotron-3.5-Lightning-30B-A3B, published as BF16 and NVFP4 repos on Hugging Face, with a NIM on build.nvidia.com. 30B total and 3B active parameters; the model card describes a hybrid of Mamba-2, MoE and a small number of attention layers.
• Licence: OpenMDW-1.1, which the model card marks as ready for commercial use. The card lists context up to 1M tokens (256K on a single H100 because of memory) and more than 20T pre-training tokens; Nvidia also released the Nemotron-RL-Agentic-Terminal-Pivot RL dataset and its post-training recipes.
• Model-card benchmarks (BF16): SWE-bench Verified 51.56, Terminal-Bench 2.1 24.58, PinchBench 85.37. Qwen 3.6 35B A3B scores 70.12, 44.38 and 88.07 on the same three; Gemma 4 26B A4B gets 57.40, 37.22 and 74.70.
• Nvidia's own claims: output speed up to 4x that of models around its size, and on PinchBench, a 10,000-task run at 86% accuracy that came in 30% quicker than Qwen3.6 35B. The corporate blog states the 30% figure against models in its class generally; the developer blog is where Qwen3.6 35B is named as the baseline.
• Weights are downloadable from Hugging Face and ModelScope, with hosted access through OpenRouter and as an NVIDIA NIM. Nvidia lists Jetson, GeForce RTX 5090 and DGX Spark as local targets plus Blackwell, Hopper and Ampere server GPUs, and says LM Studio, llama.cpp, Ollama and Unsloth support it.
• NeMo Switchyard: an open-source routing library on GitHub that hands every step in an agent's workflow to whichever model it judges most capable and efficient; developers can swap in routing algorithms that weight quality, latency or cost. Nvidia's internal benchmark claims roughly one-third the task-completion cost of using Opus 4.8 alone while holding frontier-level accuracy.

“What we're hearing is that Lightning is remarkably easy to customize.” — Kari Briski, Nvidia's vice president of generative AI, as quoted by SiliconANGLE

Why it matters

A 3B-active model under a permissive licence is something a small studio can run locally (Nvidia names the GeForce RTX 5090 as a target) or rent cheaply through OpenRouter. Nvidia is not selling it as a frontier coder: the pitch is the routine work inside long-running agents, such as tool calls, validating tool output, formatting results and delegating to subagents, while a bigger model does the planning. Its SWE-bench and Terminal-Bench scores sit well behind Qwen 3.6 35B A3B, which fits that positioning, so treat it as the cheap execution tier under a stronger planner. Switchyard is the piece that makes the split practical, and because the training data and recipes are public, tuning Lightning to your own workloads is realistic for a one-person shop. Briski told SiliconANGLE that CodeRabbit ran Nvidia's off-the-shelf AutoModel recipe for a single epoch and had a working router agent for $85 in about two hours; that is Nvidia's account, not an independent measurement.

This is an AI-written summary of the reporting credited above, read and edited by Nicholas before publishing. The facts and any quote belong to the original; the wording is ours. Read the original.