The leaked NVIDIA Nemotron 3 Nano 30B A3B, which we mentioned yesterday, has been officially released today! This is a MoE model that combines Mamba-2 and Transformer, with inference enabled or disabled. The knowledge deadline is November 28, 2025. The most impressive achievement is the AIME 25 math test score of 89.1. If this model were to use tools, it could reach 99.2 (some AIME 25 problems are easier to solve using programming methods). The LiveCodeBench (v6) programming test score is 68.3, which is considered usable. The GPQA scientific reasoning test score is 73. In comparison, the Qwen3-30B-A3B-Thinking-2507 scores 85 on AIME25, 66 on LiveCodeBench (v6), and 73.4 on GPQA. Essentially, the Nemotron-3-Nano-30B-A3B is a model optimized for mathematical ability. However, the biggest advantage of Nemotron-3-Nano-30B-A3B is that it makes both the pre-training and post-training datasets open source, making it a good choice for learning and research. Model address:
Performance parameters / 1
Performance parameters / 2
