NVIDIA Releases New Model: OpenReasoning-Nemotron

NVIDIA has launched the OpenReasoning-Nemotron model based on the Qwen 2.5 architecture. This model has demonstrated outstanding performance across multiple benchmarks in mathematics, science, and coding, surpassing o3 to become the new leader among open-source models. It is available in four parameter sizes—1.5B, 7B, 14B, and 32B—with significant score improvements especially at 7B parameters and above, reaching a top score of 78.2. Notably, the model was trained solely using supervised fine-tuning without any reinforcement learning. It exhibits strong reasoning capabilities and shows good generalization in mathematical tasks, although its performance may be weaker in multi-turn conversations and general-purpose tasks.