NVIDIA Launches Nemotron 3 Ultra: 550B Hybrid MoE for Agents

Santa Clara: In a major push to democratize frontier-grade artificial intelligence and power long-horizon enterprise agents, NVIDIA has officially introduced Nemotron 3 Ultra (550B-A55B). Built on a cutting-edge hybrid Transformer-Mamba architecture with a sparse Mixture-of-Experts (MoE) design, the open-weight foundation model features a massive 1-million-token context window and delivers up to 5x higher inference throughput compared to traditional open dense baselines. The release directly reinforces NVIDIA's full-stack strategy spanning custom silicon, CUDA kernels, and frontier open-weight models. For official open-weight checkpoints, deployment cookbooks, and framework guides, visit the NVIDIA Nemotron Developer Portal.

Architectural Innovations: Hybrid Mamba-Transformer Backbone

Unlike standard Transformers constrained by quadratic attention bottlenecks over long sequences, Nemotron 3 Ultra integrates State Space Model (SSM) Mamba-2 layers alongside standard Transformer attention layers:

• LatentMoE Sparsity (550B Total / 55B Active): Out of 550 billion total parameters, the system dynamically routes tokens through a proprietary LatentMoE mechanism to activate only 55 billion parameters per forward pass, maintaining a 10x sparsity ratio that prevents expert routing collapse.
• 1-Million-Token Context Window: Enables multi-agent workflows to preserve complete multi-turn conversation logs, persistent memory, and multi-file code repositories across multi-day jobs without re-indexing.
• Multi-Token Prediction (MTP): Generates multiple tokens in a single execution cycle, accelerating code completion, automated debugging, and terminal operations.
• Native NVFP4 on Blackwell Architecture: Optimized for NVIDIA Blackwell systems (such as GB200 and B200), native 4-bit floating-point precision enables high-density deployment with over 300 tokens per second throughput while lowering memory footprints.

Benchmark Performance and Agentic Efficiency

Rigorous evaluations across leading public AI benchmarks demonstrate Nemotron 3 Ultra's frontier reasoning and agentic problem-solving capabilities:

• Artificial Analysis Intelligence Index: Achieved a score of 48.0, establishing it as a front-runner among open-weight models.
• GPQA Diamond: Scored 86.7% in graduate-level scientific reasoning, outperforming multiple closed commercial models.
• PinchBench Productivity: Logged a 91.0% completion rate in agentic execution benchmarks.
• SWE-bench and Terminal-Bench 2.0: Solved complex software engineering and shell tasks while consuming 30% fewer tokens, sharply reducing operational compute expenses.

Multi-Teacher On-Policy Distillation (MOPD) and Licensing

NVIDIA trained the model using Multi-Teacher On-Policy Distillation (MOPD) within the open-source NeMo RL framework. Instead of a single supervision source, the student model learned simultaneously from more than 10 specialized domain models spanning mathematics, software engineering, and multi-turn planning. Alongside the weights, NVIDIA released 10 million new Supervised Fine-Tuning (SFT) samples, expanding its public repository to 50 million samples.

Distributed under the permissive NVIDIA Open Model License, enterprises can fine-tune, self-host, and embed the weights in private data centers or sovereign cloud clusters. The model features immediate support across vLLM, SGLang, TensorRT-LLM, and Hugging Face.