NVIDIA Launches Nemotron 3.5 Lightning: 30B Params with 3B Active, Built for Agent Execution

heyshrutimishra · x · 2026-08-11

NVIDIA released Nemotron 3.5 Lightning, a model optimized for the execution layer of AI agents.

Core Architecture: 30B parameter MoE model with only 3B active parameters per token, delivering 30B-level capacity at a fraction of the compute cost.
Design Philosophy: Agents spend 90% of their time on routine tasks like tool calls and result validation. Lightning handles these tasks cheaply, leaving frontier models to handle planning.
Performance: Achieves 86% accuracy on PinchBench; completes 10,000 agent tasks 30% faster than Qwen3.6 35B at similar accuracy.
Deployment: Features baked-in speculative decoding and NVFP4 quantization, allowing the same checkpoint to run across Blackwell, Hopper, and Ampere architectures.
Open Source: Fully open under OpenMDW-1.1, including weights, data, and recipes.

More from coding & agent