NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 Lets You Toggle AI Reasoning
NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 is a new large language model that packs 550 billion total parameters while activating only 55 billion during use. It combines Mamba-2, mixture-of-experts (MoE), and attention layers into a single architecture, letting it handle complex tasks like coding, math, and multi-step agent workflows. The model can read and reason over documents up to one million tokens long, and users can turn its step-by-step thinking on, off, or set it to a lighter effort level.
NVIDIA who also developed Nemotron-Labs-Diffusion-14B and Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16, trained this frontier-scale model in four stages, from pre-training on 20 trillion tokens through reinforcement learning and distillation. The company released the weights openly under a license that allows commercial and non-commercial use, giving developers a powerful tool they can run on their own hardware. It supports ten languages including English, Japanese, German, and Chinese, making it practical for global agentic applications.
Hybrid architecture and long context
Key Features
- 55B active parameters from 550B total.
- Hybrid Mamba-2, MoE, and Attention layers.
- Processes up to 1 million token contexts.
- Multi-Token Prediction for faster text generation.
- Toggle reasoning on, off, or medium effort.
- Designed for agentic and tool-use tasks.
The model fits teams building AI agents, retrieval-augmented generation systems, or chatbots that need to analyze long documents and perform tool calls. Small agencies and privacy-conscious professionals can deploy it on a node with 8 H100 or B200 GPUs, keeping sensitive data in-house. Its open license gives these users the freedom to customize and integrate the model into commercial products without negotiating proprietary terms.
Deployment requirements and future plans
You will need at least 8 high-end GPUs to run this BF16 checkpoint, with recommended setups including vLLM, SGLang, or TensorRT-LLM serving backends. At the moment TensorRT-LLM support is limited to NVIDIA Blackwell architecture such as B200 or GB200, but Hopper compatibility is on the way. All backends support chunked prefill and speculative decoding through the model’s built-in Multi-Token Prediction layers.
"Nemotron-3-Ultra-550B-A55B-BF16 is a frontier-scale large language model (LLM) trained by NVIDIA, designed to deliver strong agentic, reasoning, and conversational capabilities." — Source: Hugging Face
Project Links
