NVIDIA Nemotron 3: 1M Context, Hybrid MoE Architecture, and Open Source AI Agents
Автор: BazAI
Загружено: 2025-12-18
Просмотров: 144
The NVIDIA Nemotron 3 family of models (Nano, Super, and Ultra) is the most efficient set of open models for building high-accuracy agentic AI applications. These models excel in tasks such as reasoning, advanced math, coding, instruction following, and tool calling. The Nemotron 3 architecture uses a Hybrid Mamba-Transformer Mixture-of-Experts (MoE) design to balance speed and intelligence, providing high throughput. Crucially, Nemotron 3 models support a massive context length of up to 1 million tokens, which is key for complex multi-agent environments and long-context tasks. The currently available Nemotron 3 Nano model features 30 billion parameters with approximately 3.6 billion active parameters. The Nano model is optimized for high-throughput agentic workflows and achieves up to 3.3x higher inference throughput compared to competitive models like Qwen3-30B-A3B. Upcoming Super and Ultra models will include advanced features like LatentMoE for improved accuracy and Multi-Token Prediction (MTP) for faster text generation. NVIDIA has fully embraced transparency by releasing the model weights, training recipes, and over 3 trillion tokens of data used in training
Доступные форматы для скачивания:
Скачать видео mp4
-
Информация по загрузке:
