NVIDIA的Nemotron-Cascade 2：一款在数学和编程奥林匹克中获得金牌的30B MoE模型
📄 中文摘要
NVIDIA推出了Nemotron-Cascade 2，这是一款具有30亿参数的模型，但每个token仅激活3亿参数。尽管参数量远低于前沿模型，它在2025年国际数学奥林匹克（IMO）、国际信息学奥林匹克（IOI）和国际大学生程序设计竞赛（ICPC）世界总决赛中获得了金牌。该模型的特点包括超高效的MoE架构、Cascade RL与MOPD的创新结合、原生的逐步推理思维模式、256K token的上下文窗口，以及可通过Ollama在本地执行（需要24GB VRAM）。此外，模型的权重已在Hugging Face上开放。
📄 English Summary
NVIDIA's Nemotron-Cascade 2: A 30B MoE Model That Hits Gold Medal in Math and Coding Olympiads
NVIDIA has released Nemotron-Cascade 2, a model featuring a total of 30 billion parameters, with only 3 billion active per token. Despite having significantly fewer parameters than frontier models, it achieved gold medals at the 2025 International Mathematical Olympiad (IMO), International Olympiad in Informatics (IOI), and ICPC World Finals. Key features include an ultra-efficient MoE architecture, the innovative combination of Cascade RL and MOPD, a native step-by-step reasoning thinking mode, a context window of 256K tokens, and local execution via Ollama (requiring 24GB VRAM). The model weights are also open on Hugging Face.
Powered by Cloudflare Workers + Payload CMS + Claude 3.5
数据源: OpenAI, Google AI, DeepMind, AWS ML Blog, HuggingFace 等