NVIDIA: Nemotron 3 Ultra
🧠 AI Modelnvidia
NVIDIA's open MoE model with 55B active params, 1M context, hybrid Transformer-Mamba architecture.
Nemotron 3 Ultra is a state-of-the-art open-source model developed by NVIDIA. It employs a hybrid Transformer-Mamba MoE architecture, combining the strengths of transformer attention and Mamba state-space models for efficient long-context reasoning. With 550 billion total parameters but only 55 billion active per token, it balances performance and inference cost. The model supports a context length of 1 million tokens, enabling processing of very long documents or multi-turn conversations. It is designed for reasoning and orchestration tasks, and features include frequency penalty, logit bias, min_p, and repetition penalty. Pricing is $0.50 per million input tokens and $2.50 per million output tokens.
💡Highlights
- ├─55B active / 550B total MoE
- ├─Hybrid Transformer-Mamba
- └─1M token context length
🎯For
- ├─AI researchers
- ├─Enterprise developers
- └─Advanced LLM users