Nvidia's open Nemotron 3.5 Lightning prioritizes speed over maximum intelligence

Nvidia has launched Nemotron 3.5 Lightning, the first model of its new Nemotron 3.5 series. It is a direct successor to Nemotron 3 Nano 30B A3B and retains that model's hybrid Mamba-Transformer architecture: 31.6 billion total parameters, of which 3.6 billion are active. According to the independent benchmarking platform Artificial Analysis, the model achieves an Intelligence Index of 24, a jump of nine points over its predecessor (15). This puts Lightning on par with OpenAI's gpt-oss-120b (24) and just behind Nvidia's own Nemotron 3 Super (26), which is about four times larger. The most intelligent small models in the same class, such as Qwen3.6 35B A3B (32) or Meta's new Muse Glimmer (35), remain clearly ahead. Nvidia aims Lightning at a different point on the efficiency frontier: in preliminary tests with final NVFP4 weights, the model reaches nearly 670 tokens per second, the highest measured value among compared models and almost twice as fast as Google's Gemini 3.5 Flash-Lite (386 tokens/s). A task from the Intelligence Index thus takes about 0.5 minutes, while Qwen3.6 35B A3B takes around 3.5 minutes and Gemma 4 31B about 5.8 minutes. Proprietary models continue to dominate the overall efficiency frontier: Gemini 3.5 Flash-Lite achieves an Intelligence Index of 37 with similar time per task, GPT-5.6 Luna (max) comes to 52 points in under two minutes. The biggest improvements are seen in agentic capabilities, according to Artificial Analysis. In GDPval-AA v2, Lightning achieves an Elo score of 824, a gain of 334 points over Nemotron 3 Nano, surpassing both gpt-oss-120b (800) and the larger Nemotron 3 Super (698). In Terminal-Bench v2.1, the score rises from 7 to 24.3 percent, almost at the level of gpt-oss-120b (26.2 percent). With the permissive OpenMDW-1.1 license, Nvidia positions the model as an efficient workhorse for high-throughput agentic pipelines. According to Artificial Analysis, Nvidia worked with partners like CodeRabbit and Harvey on post-training to improve performance in specific domains. Nvidia delivers the model in BF16 and NVFP4 weights; the NVFP4 variant also scores 24 on the Intelligence Index with minimal degradation versus the higher-precision version. The reasoning model processes text only and has a context window of one million tokens. Weights are available immediately; serverless inference is offered by DeepInfra, Fireworks, FriendliAI, CoreWeave, GMI Cloud, Nebius, and Crusoe.