Nvidia Nemotron 3 Family of Models
15 hours ago
- #AI Models
- #Open Source
- #NVIDIA
- NVIDIA announces Nemotron 3, a family of open models optimized for agentic AI applications.
- The Nemotron 3 family includes three models: Nano (cost-efficient), Super (collaborative agents), and Ultra (state-of-the-art accuracy).
- Key technologies include Hybrid MoE architecture, LatentMoE, Multi-Token Prediction, NVFP4 training, and support for 1M token context length.
- Nemotron 3 Nano outperforms GPT-OSS-20B and Qwen3-30B-A3B on benchmarks with higher throughput and supports 1M token context.
- Open-source releases include model checkpoints, datasets (Nemotron-CC-v2.1, Nemotron-CC-Code-v1), and model recipes.
- Datasets feature curated and synthetic data for pretraining, specialized areas (STEM, coding), and reinforcement learning.