NVIDIA Releases Nemotron 3 Ultra: 550B Open-Weight LLM with Industry-Leading Inference Performance
Key Takeaways
- ▸Nemotron 3 Ultra (550B/55B) delivers 1.6–5.9x higher inference throughput than competing open-weight LLMs, establishing a new performance bar for efficient large-scale models
- ▸Supports 1M token context length with state-of-the-art performance, enabling long-document and extended reasoning use cases without quality degradation
- ▸Full open-source release includes model weights, 173B tokens of fresh training code, synthetic specialized datasets, and pretraining code—significantly lowering barriers to frontier model development
Summary
NVIDIA unveiled Nemotron 3 Ultra, its most capable open-weight language model to date, featuring 550 billion total parameters and 55 billion active parameters through a Mixture-of-Experts (MoE) Hybrid Mamba-Attention architecture. The model achieves competitive accuracy across industry benchmarks while delivering significant performance advantages—5.9x higher inference throughput than GLM-5.1-754B, 4.8x over Kimi-K2.6-1T, and 1.6x over Qwen-3.5-397B on standard inference configurations.
Nemotron 3 Ultra supports up to 1 million token context length while maintaining state-of-the-art performance on RULER benchmarks, making it suitable for long-document processing and complex reasoning tasks. The model incorporates LatentMoE for accuracy improvements and MTP layers for faster inference through native speculative decoding, plus inference-time reasoning budget control for flexible deployment.
Beyond the model weights, NVIDIA is releasing a comprehensive open-source ecosystem: pre-trained, post-trained, and quantized checkpoints (NVFP4 and BF16 formats); 173B tokens of fresh code data from GitHub through September 2025; synthetic datasets for legal and specialized domain capabilities; and the pretraining code itself. The post-training methodology—combining Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD)—is fully documented in the accompanying tech report.
- Advanced training pipeline (SFT + RL + MOPD) and architectural innovations (Hybrid Mamba-Attention, LatentMoE, speculative decoding) combined in a single production-ready model
Editorial Opinion
Nemotron 3 Ultra marks a watershed moment in democratizing large-scale LLMs. By releasing not just weights but training datasets, code, and synthetic domain datasets, NVIDIA is fundamentally lowering the cost and complexity of competitive model development. The combination of efficient inference, million-token context, and strong accuracy makes this a credible open-source alternative to proprietary frontiers—and signals that the performance gap between open and closed models is narrowing at scale.
