After Huang Renxun supported open source AI, NVIDIA launched two major open source weapons! The output speed is 4 times faster and the cost is cut to 1/3

Compiled by Zhidongzhi | Edited by Eggplant | Cheng Qian Zhidongzhi reported on August 12 that yesterday, NVIDIA released the open source large language model Nemotron 3.5 Lightning, and the open source intelligent routing library NeMo Switchyard for Agents. The former has a total parameter volume of 30 billion, but only about 3 billion parameters are activated during inference, and the output speed is up to 4 times that of a model of the same size; the latter can automatically allocate tasks among multiple models. NVIDIA's internal tests show that its task completion cost can be reduced to 1/3 of that when using Claude Opus 4.8. ▲Nemotron 3.5 Lightning and NeMo Switchyard are open source (Source: Hugging Face, GitHub) Nemotron 3.5 Lightning is the first open source model launched by Nvidia after signing a joint open letter supporting open source AI models on July 24. Nvidia founder and CEO Jensen Huang said that Nemotron 3.5 Lightning is suitable for agents that run continuously and work for a long time, and summarized it as "smart, fast, efficient and open source." ▲Huang Renxun forwarded the news of Nemotron 3.5 Lightning release (Source: X) Artificial Analysis, a large model evaluation agency, compared open weight models with a total parameter volume of less than 40 billion. Nemotron 3.5 Lightning has a comprehensive ability score of 24 points and an output speed of approximately 670 token/s. It is the only model in the figure that enters the high-speed and high-scoring area. ▲Evaluation results of Nemotron 3.5 Lightning in Artificial Analysis (Source: in technology