NVIDIA Nemotron 3.5 Lightning: The High-Speed Execution Layer for the Agentic Era

Introduction: The New Standard for Agentic Execution

On August 11, 2026, NVIDIA officially shifted the landscape of open-source AI with the release of Nemotron 3.5 Lightning. This isn't just another incremental update; it is a purpose-built powerhouse designed to solve the 'orchestrator-executor' dilemma. While frontier models like Nemotron 3 Ultra act as the high-level brain, Lightning is engineered to be the high-volume execution layer—the muscle that handles the relentless, high-frequency tasks required by autonomous agents.

• First release in the Nemotron 3.5 family
• Successor to Nemotron 3 Nano 30B A3B
• Released under the permissive OpenMDW-1.1 license
• Designed for high-volume, always-on agentic workflows

Architecture: Efficiency Meets Intelligence

At the heart of Nemotron 3.5 Lightning is a sophisticated Mixture-of-Experts (MoE) architecture. The model boasts a total of 31.6B parameters, but thanks to its MoE design, only 3.6B parameters are active during any single forward pass. This allows for massive intelligence density without the massive computational overhead typically associated with large-scale models.

• Total Parameters: 31.6B
• Active Parameters: 3.6B (MoE)
• Hybrid Mamba-Transformer architecture
• 1 Million token context window
• Text-only reasoning optimization

Performance: Breaking the Speed-Intelligence Barrier

The performance leap from the previous Nemotron 3 Nano is staggering. On the Artificial Analysis Intelligence Index, Nemotron 3.5 Lightning scored a 24, representing a +9 point jump from its predecessor. To put that in perspective, it matches OpenAI's gpt-oss-120b while utilizing only about a quarter of the total parameters. It trails only the massive Nemotron 3 Super (26), which is roughly 4x its size.

• Intelligence Index: 24 (vs 15 for Nemotron 3 Nano)
• Terminal-Bench v2.1: Jumped from 7% to 24%
• GDPval-AA v2: +334 ELO gain (outperforming gpt-oss-120b)
• Speed: Nearly 670 tokens per second (via DeepInfra NVFP4 endpoint)

Quantization and Deployment: The NVFP4 Advantage

NVIDIA has optimized the deployment experience by shipping the model with near-lossless NVFP4 quantization. This isn't just a compression trick; the NVFP4 weights themselves measured at a 24 on the Intelligence Index, proving that precision is maintained even at extreme efficiency. This makes the model highly accessible for edge deployment or single-GPU setups.