According to CoinMeta, NVIDIA has released the open-source model Nemotron 3.5 Lightning, which is specifically designed for tool calls in long-running Agent. This model has a total of 30 billion parameters, with only 3 billion parameters activated per token, and is mainly responsible for tool calls, result checks, and scheduling of sub-tasks such as Agent. NVIDIA's approach is to divide the workload more thoroughly; complex planning is entrusted to large models like Nemotron 3 Ultra, while a large number of repetitive tasks are handled by Lightning. Officials claim that its output speed can reach up to four times that of models of the same level. Based on artificial intelligence analysis, Nemotron 3.5 Lightning performed exceptionally well in independent evaluations, achieving a speed of 416 token per second, ranking first among open-weight models of the same size. Although its intelligence index is 24 points, it showed significant improvement in Agent capabilities, reaching 824, surpassing gpt-oss-120b and Nemotron 3 Super.