According to CoinMeta, NVIDIA officially open-sourced its flagship large language model with 55 billion parameters, Nemotron 3 Ultra, on June 4th. This model adopts a pioneering Mamba-Transformer hybrid MoE architecture that is specifically optimized for long-term agent tasks. On third-party benchmark platforms, this model scored 47.7 points in the intelligence index, making it the most powerful open-source language model in the United States. NVIDIA also plans to launch Nemotron 4, with parameters expected to reach 1 trillion. The specific specifications and release time are not yet determined. The company hopes to drive demand for GPU through more powerful open models, and its cloud service procurement commitments for the coming years have increased to 28 billion US dollars, which is approximately three times that of a year ago.