#AGI Nvidia 開源 Nemotron-Nano-9B-v2 模型，支援可切換「思考模式」，並在多個基準測試中擊敗阿里巴巴的 Qwen3-8B。模型結構融合 Transformer 與 Mamba，能在單張 A10 GPU 上運行，並於 MATH500 測試中達 97.8%，在 IFEval 指令跟隨評估中更達 90.3%，全面領先 Qwen3-8B 的 92.4% 及 84.9%。支援多語言、程式生成與開源商用，展現 Nvidia 推進推理透明度與小模型效能的企圖。
Nvidia has launched Nemotron-Nano-9B-v2, a compact yet powerful open-source LLM featuring toggleable reasoning. It outperforms Alibaba’s Qwen3-8B in key benchmarks—scoring 97.8% vs 92.4% on MATH500, and 90.3% vs 84.9% on IFEval. Built with hybrid Transformer-Mamba layers, it runs on a single A10 GPU, supports multilingual and code tasks, and is licensed for free commercial use. A major step forward for efficient, controllable AI.
📌 一杯咖啡價錢連接 Web3 世界 [https://patreon.com/wanszezit](https://patreon.com/wanszezit)
Full article https://venturebeat.com/ai/nvidias-open-nemotron-nano-9b-v2-has-toggle-on-off-reasoning/