Nvidia's Nemotron 3.5 Lightning matches larger models at quarter the size

What happened

Nvidia released Nemotron 3.5 Lightning, a compact open-weights model with 31.6 billion total parameters (only 3.6 billion active per step) that scores 24 on the Artificial Analysis Intelligence Index—matching OpenAI's gpt-oss-120b—while delivering 669 tokens per second, the fastest inference speed in its class.

Why it matters

The model demonstrates that smaller, efficient architectures can match larger competitors on reasoning tasks. For developers building agent-based systems, Lightning offers a permissively-licensed alternative (OpenMDW-1.1) that runs nearly twice as fast as Google's Gemini 3.5 Flash-Lite (386 tokens/s) while achieving parity on intelligence benchmarks, reducing computational cost and latency.

What to watch

Lightning is available now in both BF16 and NVFP4 weights, with serverless inference offered by DeepInfra, Fireworks, FriendliAI, CoreWeave, GMI Cloud, Nebius, and Crusoe. The model shows the biggest gains on agentic benchmarks, scoring an Elo rating of 824 on GDPval-AA v2 and 24.3% on Terminal-Bench v2.1.