Nemotron-3-Ultra-550B-A55B-NVFP4

Key Features

Read the Model Signal report →

Identity

Nemotron-3-Ultra-550B-A55B-NVFP4 is a frontier-scale general-purpose AI model developed by NVIDIA, released on June 4, 2026 (model profile). It is a hybrid model utilizing a LatentMoE Mamba-2 and Mixture-of-Experts (MoE) architecture, designed for high-parameter efficiency and performance using the NVFP4 precision format (model profile).

What it is

The model is a large-scale system featuring 550 billion total parameters, with 55 billion active parameters per token (model profile). Its architecture combines LatentMoE Mamba-2 with MoE components (model profile). It is designed to handle extensive context, though reported capacities vary between 512,288 and 1,000,000 tokens (model profile). The model utilizes NVIDIA's NVFP4 precision format for its operations (model profile).

Capabilities & benchmarks

The following specifications are reported by the provider (model profile):

• Provider: NVIDIA
• Release Date: June 4, 2026
• Context Window: 512,288 tokens (structured data) / 1,000,000 tokens (summary/key features)
• Category: General
• Architecture: Hybrid LatentMoE Mamba-2 + MoE
• Parameters: 550 billion total; 55 billion active per token
• Precision: NVFP4 format

Where it fits

The model is categorized as a general-purpose, frontier-scale AI model (model profile). Its high parameter count and hybrid architecture suggest it is intended for complex, large-scale reasoning and long-context tasks (model profile).

Contradictions

• Context Window Discrepancy: The model profile contains conflicting data regarding the context window size. The structured metadata field lists 512,288 tokens, while the summary and key features sections both claim a 1,000,000 token context window (model profile).

Sources

• model profile: NVIDIA model metadata for Nemotron-3-Ultra-550B-A55B-NVFP4.
• provider announcement: NVIDIA official model card and announcement.