Nvidia has launched its new Llama-3.1 Nemotron Ultra, a 253 billion parameter model that outperforms the DeepSeek R1, which has 671 billion parameters, while being less than half its size. The new model boasts four times higher inference throughput compared to DeepSeek R1 and achieves the highest accuracy on several reasoning benchmarks, including GPQA-Diamond and AIME 2024/25. In addition, the open-source company Deep Cogito has released a suite of language models ranging from 3 billion to 70 billion parameters, each outperforming their counterparts from other models, including LLaMA and DeepSeek. These developments highlight the competitive landscape in artificial intelligence, particularly in the realm of large language models, as companies strive to enhance performance and efficiency.