Nvidia's Nemotron 3.5 Lightning bets on speed over raw IQ

Nvidia has released Nemotron 3.5 Lightning, an open-weights language model built around a simple trade-off: outrun the competition rather than out-think it. The model packs 31.6 billion total parameters but activates only 3.6 billion at a time, a mixture-of-experts style design that keeps inference cheap while still leaving room for capable reasoning.

The headline number is throughput. Nemotron 3.5 Lightning hits nearly 670 tokens per second, almost double the 386 tokens per second Nvidia clocked for Google's Gemini 3.5 Flash-Lite in the same comparison. That makes it the fastest model in Nvidia's benchmark set by a wide margin.

Speed doesn't come entirely at the expense of intelligence, either. The model scores 24 on the Artificial Analysis Intelligence Index, matching OpenAI's gpt-oss-120b despite having roughly a quarter of its parameter count. That's a nine-point jump over its predecessor, Nemotron 3 Nano 30B, which scored 15.

On Terminal-Bench v2.1, a benchmark for agentic coding tasks, Nemotron 3.5 Lightning posts 24.3 percent, close behind gpt-oss-120b's 26.2 percent. Its gains look even sharper on GDPval-AA v2, an agentic evaluation where it reaches an Elo rating of 824, up 334 points from the prior Nemotron generation.

Some of that jump traces back to targeted post-training work. Nvidia says it partnered with CodeRabbit and Harvey, companies focused on code review and legal AI respectively, to sharpen the model's performance in those specific domains rather than chasing generic benchmark gains alone.

Under the hood, the model uses a hybrid Mamba-Transformer architecture, pairing the efficiency of state-space models with the pattern-matching strength of standard attention layers. It's text-only, but supports a one-million-token context window, putting long-document and long-conversation use cases within reach.

Nvidia is shipping the model under the permissive OpenMDW-1.1 license, and in two weight formats: standard BF16 and the compressed NVFP4 format. Notably, the NVFP4 variant still scores 24 on the Intelligence Index, suggesting the aggressive quantization costs little in practical quality.

Nemotron 3.5 Lightning prioritizes speed over maximum intelligence, matching a model four times its size while running nearly twice as fast as Google's comparable offering.

For teams weighing inference costs against model capability, Nemotron 3.5 Lightning positions itself as a middle path: not the smartest open model available, but one engineered specifically to be fast and cheap enough to run at scale without a steep accuracy penalty.