What hardware does Nemotron 3.5 Lightning run on?

Nemotron 3.5 Lightning is built to run either locally or in the cloud. NVIDIA lists RTX PCs, DGX Spark, DGX Station and Jetson as local targets, and the same weights are served in the cloud through build.nvidia.com as an NIM microservice and through NVIDIA Cloud Partners, OpenRouter and ModelScope.

Which companies already use Nemotron 3.5 Lightning?

NVIDIA names five early adopters for Nemotron 3.5 Lightning in its launch post: CrowdStrike in cybersecurity, Harvey with Trajectory in legal work, CodeRabbit with Baseten for code review, Lila Sciences in physical and life sciences, and Fastino Labs across software, finance and healthcare.

What languages does Nemotron 3.5 Lightning support?

The Nemotron 3.5 Lightning model card lists English plus coding languages, Spanish, French, German, Italian and Japanese. NVIDIA measured its published accuracy numbers under one consistent harness, NeMo Gym and the NeMo Evaluator SDK, and notes they may differ from numbers other vendors report for their own models.

Can I fine-tune Nemotron 3.5 Lightning on my own data?

Yes. NVIDIA describes Nemotron 3.5 Lightning as a fully customizable open model and says teams can post-train it with NVIDIA NeMo on their own organization-specific data. The weights ship under the permissive OpenMDW-1.1 license in both NVFP4 and BF16 checkpoints, so you pick the precision your hardware fits.

How fast is Nemotron 3.5 Lightning compared to models its size?

NVIDIA claims Nemotron 3.5 Lightning delivers up to 4x the output speed of similar-sized models and finishes agentic tasks 30% faster. On PinchBench it reaches 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy, and NVIDIA says it wins the accuracy-versus-speed Pareto frontier on the Artificial Analysis Intelligence Index.

NVIDIA · 2026-08-11 · major

Nemotron 3.5 Lightning — NVIDIA's 30B open MoE for always-on agents

Nemotron 3.5 Lightning is a 30B open mixture-of-experts model with 3B active parameters, released by NVIDIA on 2026-08-11. It handles up to 1M tokens of context and targets high-volume agent steps that run all day.

NVIDIA Nemotron 3.5 Lightning agentic AI announcement graphic

NVIDIA's new 30B open MoE keeps only 3B parameters active per token, aimed at agents that fire thousands of small steps.

Key specs

SWE-bench	52.8%
GPQA	75.57%

Quick facts

Maker	NVIDIA
Parameters	30B total, 3B active
Architecture	Mamba-2 + MoE + attention hybrid
Context window	Up to 1M tokens
License	OpenMDW-1.1
Checkpoints	NVFP4 and BF16
Availability	Hugging Face, ModelScope, OpenRouter, build.nvidia.com NIM

What is it?

Nemotron 3.5 Lightning is a new open-weight model NVIDIA released on 2026-08-11 for the cheap, repetitive steps inside a long-running agent. It holds 30B parameters but activates only 3B of them per token, reads up to 1M tokens of context, and ships under the permissive OpenMDW-1.1 license in NVFP4 and BF16 checkpoints.

How does it work?

The architecture is a hybrid: Mamba-2 state-space layers sit alongside attention layers, with a mixture-of-experts router picking a small slice of experts per token. That combination is what keeps the active parameter count at 3B, which is where the speed comes from — up to 4x the output rate of similar-sized models, and 30% faster completion on agentic tasks.

Why does it matter?

Agent workloads are dominated by small steps, and paying frontier prices for each one adds up. A 3B-active model that scores 52.8% on SWE-bench Verified and 81.62 on MMLU Pro lets teams move that volume onto an RTX PC, a Jetson board, or a cheap cloud endpoint. CrowdStrike, Harvey, CodeRabbit, Lila Sciences and Fastino Labs are already building on it.

Who is it for?

agent developers, teams running high-volume LLM workloads

Frequently asked questions

What hardware does Nemotron 3.5 Lightning run on?: Nemotron 3.5 Lightning is built to run either locally or in the cloud. NVIDIA lists RTX PCs, DGX Spark, DGX Station and Jetson as local targets, and the same weights are served in the cloud through build.nvidia.com as an NIM microservice and through NVIDIA Cloud Partners, OpenRouter and ModelScope.
Which companies already use Nemotron 3.5 Lightning?: NVIDIA names five early adopters for Nemotron 3.5 Lightning in its launch post: CrowdStrike in cybersecurity, Harvey with Trajectory in legal work, CodeRabbit with Baseten for code review, Lila Sciences in physical and life sciences, and Fastino Labs across software, finance and healthcare.
What languages does Nemotron 3.5 Lightning support?: The Nemotron 3.5 Lightning model card lists English plus coding languages, Spanish, French, German, Italian and Japanese. NVIDIA measured its published accuracy numbers under one consistent harness, NeMo Gym and the NeMo Evaluator SDK, and notes they may differ from numbers other vendors report for their own models.
Can I fine-tune Nemotron 3.5 Lightning on my own data?: Yes. NVIDIA describes Nemotron 3.5 Lightning as a fully customizable open model and says teams can post-train it with NVIDIA NeMo on their own organization-specific data. The weights ship under the permissive OpenMDW-1.1 license in both NVFP4 and BF16 checkpoints, so you pick the precision your hardware fits.
How fast is Nemotron 3.5 Lightning compared to models its size?: NVIDIA claims Nemotron 3.5 Lightning delivers up to 4x the output speed of similar-sized models and finishes agentic tasks 30% faster. On PinchBench it reaches 86% accuracy while completing 10,000 tasks 30% faster than Qwen3.6 35B at similar accuracy, and NVIDIA says it wins the accuracy-versus-speed Pareto frontier on the Artificial Analysis Intelligence Index.

Try it

Pull `nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4` from Hugging Face