NVIDIA Unveils Nemotron 3.5 Lightning: A New Generation of Efficient AI
NVIDIA has announced Nemotron 3.5 Lightning, a new language model purpose-built for long-running AI agents and always-on agentic workflows. With 30 billion total parameters but only 3 billion active at any time, the model strikes a notable balance between speed and accuracy — a combination that's becoming increasingly critical as AI agents move from experimental demos into production systems handling high volumes of tool calls, validations, and task delegation.
What's Under the Hood
Nemotron 3.5 Lightning uses a hybrid Mixture-of-Experts (MoE) architecture, interleaving Mamba-2 and MoE layers with select Attention layers. This design lets the model activate only a fraction of its total parameters per token, delivering strong reasoning capability without the computational overhead of a fully dense model of similar scale.
The model supports a 1 million token context window and was trained on over 20 trillion tokens. It ships in both BF16 and NVFP4 quantized formats — the latter using the same specialized kernels that power NVIDIA's larger Nemotron 3 Ultra model, running efficiently across Blackwell, Hopper, and Ampere GPU generations. Notably, NVIDIA reports minimal accuracy loss between the full-precision and NVFP4 versions, making the smaller footprint practical for real-world deployment.
Built for Speed
A standout feature is the model's use of speculative decoding, paired with dedicated companion checkpoints (like the DSpark and DFlash variants) that accelerate token generation. NVIDIA claims up to 4x faster output speed compared to similarly sized models — a meaningful advantage for agentic applications where an AI system might be making dozens of rapid tool calls in sequence rather than generating a single long response.
Where It Fits
According to independent benchmarking from Artificial Analysis, Nemotron 3.5 Lightning scores 24 on their Intelligence Index — a 9-point jump over its predecessor, Nemotron 3 Nano — putting it roughly on par with OpenAI's gpt-oss-120b despite having a fraction of the parameters. It sits just behind Nemotron 3 Super, a model about four times its size.
The model is designed to run efficiently not just in data centers, but also on local hardware including NVIDIA Jetson, GeForce RTX 5090, and DGX Spark, putting capable agentic AI within reach of individual developers and smaller teams.
Licensing and Availability
Nemotron 3.5 Lightning is released under NVIDIA's OpenMDW-1.1 license, permitting commercial use without significant restrictions. Weights are available now on Hugging Face, alongside serverless inference support from providers including DeepInfra, Fireworks AI, Friendli AI, CoreWeave, GMI Cloud, and Nebius.
