NVIDIA Nemotron 3 Ultra 550B: Open-Weight MoE Model Built for Long-Horizon Agents

NVIDIA Joins the Open-Weight Frontier

On June 4, 2026, NVIDIA released Nemotron 3 Ultra — a fully open-weight large language model with 550 billion total parameters (55 billion active per token). Unlike NVIDIA's previous Nemotron releases, which were primarily fine-tuned derivatives of Meta's Llama models, Nemotron 3 Ultra is a ground-up architecture designed specifically for long-running agentic workloads. The weights, training data, and evaluation recipes are all publicly available under the OpenMDW-1.1 license, making this one of the most capable openly licensed models to date.

The release positions NVIDIA not just as a hardware vendor but as a first-class contributor to the open AI research ecosystem — a shift with broad implications for enterprise deployments, academic research, and the competitive dynamics between open and closed frontier models.

Architecture: Hybrid Mamba-Transformer MoE

The defining technical characteristic of Nemotron 3 Ultra is its hybrid LatentMoE architecture, which combines three distinct layer types:

Mamba-2 layers handle the bulk of sequence processing using a state-space model that scales sub-quadratically with context length. This means that processing a 1-million-token context does not incur the quadratic memory and compute costs that plague standard Transformer attention.

Selective Attention layers are interspersed at regular intervals to preserve precise recall over specific passages within the long context. Where Mamba layers excel at broad pattern recognition, Attention layers anchor the model to exact tokens when precision matters.

LatentMoE routing governs which of 512 experts per layer are activated for any given token. Only the top 22 experts fire per token, keeping active parameters at 55B despite the 550B total parameter count.

The result is a model that NVIDIA benchmarks at up to 6x higher inference throughput than comparable open LLMs, with native 1M-token context achieved without retrieval augmentation or context compression tricks.

Multi-Token Prediction and Speculative Decoding

The checkpoint ships with two Multi-Token Prediction (MTP) layers baked in, enabling native speculative decoding at 300+ tokens per second without any external draft model. NVIDIA refers to this as "built-in speculative decoding" — the MTP layers act as a lightweight draft head that the main model can verify in parallel, squeezing extra throughput out of Blackwell, Hopper, and Ampere GPUs without additional inference infrastructure.

Training Recipe

Nemotron 3 Ultra was pre-trained on approximately 20 trillion tokens. The post-training stack draws on Multi-teacher On-Policy Distillation from more than 10 specialized teacher models, a technique NVIDIA calls "compound distillation." Rather than relying on a single teacher, this approach pulls signal from models that excel at different tasks — coding, reasoning, long-context retrieval, instruction following — and blends those signals during reinforcement fine-tuning.

Benchmark Performance

The SWE-Bench Verified score of 71.9 is particularly notable. This benchmark measures a model's ability to autonomously resolve real GitHub issues, and a score above 70 is generally considered "frontier-competitive." Among open-weight models, Nemotron 3 Ultra is among the first US-developed models to cross that threshold.

The RULER score of 94.7 at 1 million tokens means the model reliably retrieves information from arbitrary positions across a 1M-token context — a capability critical for the long-running agent use cases NVIDIA targets.

Deployment Options

Nemotron 3 Ultra ships through multiple channels as of June 4, 2026:

• HuggingFace: Base (BF16), instruction-tuned (BF16), GenRM reward model, and NVFP4 quantized checkpoints
• NVIDIA NIM: Containerized microservice for enterprise deployment
• OpenRouter: Available for API access
• Together AI: Hosted inference
• Nebius: Cloud deployment
• Perplexity: Research access

A community GGUF quantization is also available on HuggingFace from Unsloth, as well as a 4-bit MLX version for Apple Silicon.

The inference-time budget control feature deserves special mention: a medium-effort reasoning mode trades approximately 7% accuracy for a 2.5x reduction in token generation, allowing cost-sensitive deployments to dial back compute when maximum accuracy is not required.

Usability Analysis

For most organizations, Nemotron 3 Ultra sits in the "powerful but infrastructure-heavy" tier. The full BF16 checkpoint requires multi-GPU deployment, with the NVFP4 checkpoint being the practical starting point for most hardware — compatible with Blackwell, Hopper, and Ampere GPUs.

The NIM containerization dramatically lowers the operational barrier for enterprise teams already in the NVIDIA ecosystem. For those already running NVIDIA-accelerated infrastructure, deploying Nemotron 3 Ultra via NIM is comparable in complexity to deploying any other NIM-packaged model.

Researchers working in the Hugging Face ecosystem will find the BF16 base and instruct checkpoints straightforward to integrate with standard transformers or vLLM pipelines. The GGUF version from Unsloth also opens access to CPU-based and mixed CPU/GPU inference for smaller labs.

The 1M-token context with strong RULER scores makes this a serious option for document analysis pipelines, codebase-level agent tasks, and multi-step research workflows that require retaining context across many tool calls.

Pros and Cons

Strengths:

• First US open-weight model to combine 550B parameters, 1M context, and agentic SWE-bench scores above 70
• OpenMDW-1.1 license permits commercial use with clear terms
• Built-in speculative decoding via MTP layers reduces inference cost without a separate draft model
• Broad deployment support: HuggingFace, NIM, OpenRouter, Together AI
• Inference-time budget control for cost-performance tuning

Limitations:

• Full BF16 weights require substantial multi-GPU infrastructure
• Artificial Analysis Index of 48 places it second among US open models — the gap to China's Kimi K2.6 on the overall intelligence index is acknowledged in NVIDIA's own documentation
• OpenMDW-1.1 is a new license not yet as widely understood or vetted by legal teams as Apache 2.0 or MIT
• Long-context performance requires Blackwell/Hopper GPUs for practical deployment at 1M token lengths

Outlook

Nemotron 3 Ultra represents a meaningful escalation in NVIDIA's open AI strategy. The company now offers a complete stack: hardware (Blackwell GPUs), inference optimization (NIM, NVFP4 quantization, built-in speculative decoding), and — as of June 4 — a competitive frontier model. This vertical integration has a clear competitive logic: every organization that standardizes on Nemotron 3 Ultra for inference is also an organization that stays in the NVIDIA GPU ecosystem.

For the broader open-source AI ecosystem, Nemotron 3 Ultra raises the benchmark for what open-weight models can achieve. If NVIDIA continues to release models at this capability tier under permissive licenses, the practical performance gap between open and closed frontier models will continue to narrow — which benefits research reproducibility, enterprise risk management, and competitive diversity.

Conclusion

NVIDIA Nemotron 3 Ultra 550B is the most technically sophisticated open-weight model NVIDIA has released to date. Its hybrid Mamba-Transformer MoE architecture solves real engineering problems — long-context efficiency and inference throughput — rather than simply scaling parameter count. For AI engineering teams building long-horizon agents, multi-step reasoning pipelines, or large-codebase analysis tools, Nemotron 3 Ultra is now the leading open-weight option in the US ecosystem. Hardware requirements are significant, but the combination of NIM deployment, NVFP4 quantization, and built-in speculative decoding makes production deployment tractable for well-resourced teams.

Editor's Verdict

NVIDIA Nemotron 3 Ultra 550B: Open-Weight MoE Model Built for Long-Horizon Agents earns a solid recommendation within the open source space.

The strongest case for paying attention is best-in-class long-context recall at 1M tokens (RULER 94.7) with sub-quadratic memory scaling, which raises the bar for what readers should now expect from peers in this space. Reinforcing that, commercial-use permissive licensing under OpenMDW-1.1 with full transparency adds practical value rather than just headline appeal. The broader signal worth registering is straightforward: nemotron 3 Ultra is NVIDIA's first ground-up LLM architecture — not a Llama derivative — signaling a strategic shift to becoming a full-stack AI model provider. On the other side of the ledger, full-precision BF16 deployment requires substantial multi-GPU infrastructure not accessible to small teams is a real constraint, not a marketing footnote, and it should factor into any serious decision. Layered on top of that, openMDW-1.1 is a novel license that legal teams at risk-averse enterprises may need time to review and approve narrows the set of teams for whom this is an obvious yes.

For developers building locally, infrastructure engineers, and anyone preferring transparent, modifiable software, this is a serious evaluation candidate, not just a curiosity to bookmark. For everyone else, the safer posture is to monitor coverage and revisit once the use cases that matter to your team are demonstrated in the wild.