[NVIDIA Nemotron 3.5 Lightning Real-World Test: Intelligence Not Top-Tier, Speed Takes the Lead] According to monitoring by Beating, Artificial Analysis has completed an independent evaluation of Nemotron 3.5 Lightning. The model has only 31.6 billion total parameters and 3.6 billion active parameters, achieving an intelligence score of 24, on par with OpenAI's gpt-oss-120b, which has approximately four times the parameters. This score is 9 points higher than the previous generation Nemotron 3 Nano. Lightning's standout feature is still its speed. Pre-release tests conducted by Artificial Analysis on DeepInfra measured speeds close to 670 tokens per second; the official leaderboard currently lists it at 416 tokens per second, ranking first among 134 comparable open-weight models of the same size. Agent capabilities have also seen significant improvement. On GDPval-AA v2, its Elo score reached 824, surpassing gpt-oss-120b and Nemotron 3 Super. On Terminal-Bench v2.1, its performance increased from 7% in the previous generation to 24%, a more than threefold improvement. However, it is not the smartest model of its size. On Artificial Analysis's intelligence index, Qwen3.6 35B A3B scored 32 points, and Muse Glimmer High scored 35 points, while Lightning only scored 24 points. NVIDIA did not aim for the highest intelligence score in its size category but instead focused on speed and Agent execution. [Original Link]