Nemotron 3 Ultra: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

Abstract: We introduce Nemotron 3 Ultra, a 550 billion total and 55 billion active parameter Mixture-of-Experts Hybrid Mamba-Attention LLM. We pre-trained Nemotron 3 Ultra on 20 trillion text tokens, then extended the context length to 1M tokens, and post-trained using Supervised Fine Tuning (SFT), Reinforcement Learning (RL), and Multi-teacher On-Policy Distillation (MOPD). Nemotron 3 Ultra is our most capable model yet, employing multiple key technologies - LatentMoE, Multi Token Prediction (MTP), NVFP4 pre-training, multi-environment RLVR, MOPD, and reasoning budget control. Nemotron 3 Ultra achieves up to ~6x higher inference throughput as compared to state-of-the-art publicly available LLMs while attaining on-par accuracy. The state-of-the-art accuracy, high inference throughput, and 1M token context length make Nemotron 3 Ultra ideal for long-running autonomous agentic tasks. We open-source the base, post-trained, and quantized checkpoints, along with the training data and recipe on HuggingFace.

• The paper introduces Nemotron 3 Ultra, a 550B-parameter open MoE LLM leveraging a hybrid Mamba-Attention architecture optimized for long-context agentic reasoning.
• It employs advanced LatentMoE routing and multi-token prediction to enhance sparse expert capacity, boosting inference throughput up to 1.6× over competitors.
• The integrated post-training pipeline with MOPD and RLVR significantly improves agent performance, ethical decision-making, and inference stability.

Nemotron 3 Ultra: An Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning

Nemotron 3 Ultra (also referred to as Nemotron Nano 3 in the text) represents a significant large-scale advance in open Mixture-of-Experts (MoE) language modeling optimized for agentic, long-context, and efficient inference scenarios. The model combines an immense scale, hybrid Mamba-Attention backbone, advanced sparsity via LatentMoE, and an intricate post-training pipeline built for robust agentic decision making, tool use, and verifiable reasoning.

---

Model Architecture and Pretraining

Nemotron 3 Ultra leverages a 550B total parameter (55B active per token) MoE architecture that positions it at the vanguard of inference-aware design for open LLMs.

The core architectural choices include:

• Hybrid Mamba-Attention Backbone: Adopts Mamba state space layers interleaved with sparse global attention, providing linear-time scaling in decoding, strong long-context handling (up to 1M tokens), and reduced KV cache footprint.
• LatentMoE Routing: Employs LatentMoE to maximize expert capacity per FLOP/cost, activating 22 out of 512 experts per MoE layer, scaling sparsely, and trading hidden dimension width for number of experts.
• Multi-Token Prediction (MTP): Native support for speculative decoding to accelerate inference, with dedicated MTP heads tightly integrated (Figure 1).

The pretraining phase encompasses 20 trillion text tokens sourced from a high-quality and extensively filtered blend. Key innovations in data include very large-scale code refreshes, injection of synthetic multi-choice and generative QA spanning diverse domains, fact-seeking and moral scenarios, and sophisticated legal domain datasets. The data mixture is dynamically controlled via a two-phase curriculum focusing on diversity (phase 1) and quality (phase 2) (Figure 2).

Training is performed primarily in NVIDIA's NVFP4 precision, with selective high-precision retention in the final layers and sensitive projections, marking the largest stable NVFP4 training to date. Abrupt loss divergences were encountered and extensively characterized, leading to early termination at 20T tokens, demonstrating the persistent challenges in ultra-scale pretraining with sparse architectures.

Figure 1: Nemotron Nano 3 layer pattern employing a hybrid Mamba-Attention architecture sparsely scaled via LatentMoE layers.

Post-Training: SFT, RLVR, and MOPD

Supervised Fine-Tuning and RL

The post-training pipeline is designed for agentic and tool-use environments, extending beyond classic instruction SFT and RLHF recipes. After large-scale SFT focused on agentic, reasoning, safety, search, tool-use, code, math, and multilingual scenarios, Nemotron 3 Ultra undergoes unified RLVR (Reinforcement Learning with Verifiable Reward) across an expansive set of environments.

Multi-teacher On-Policy Distillation (MOPD)

A central innovation is an asynchronous, multi-iteration MOPD protocol, in which over ten domain-specialized teacher models (optimized for domains such as SWE, GDPVal, search, terminal use, safety, etc.) provide dense, token-level distillation signals on student rollouts. This iterative student–teacher co-evolution bridges the gap between generalist student and domain-optimized specialists while avoiding instability from multitask RL.

Figure 3: Two-iteration MOPD training pipeline for Nemotron Nano 3, showing iterative specialization and distillation from multiple teachers.

Ablations demonstrate MOPD closes the teacher–student gap by 70–90% in agentic domains; limitations appear primarily in single-turn reasoning domains where supervision via teacher disagreement is less effective. Warm-up SFT on the teacher's data distribution before MOPD measurably improves performance by mitigating distribution mismatch.

---

Quantization and Inference Optimization

Mixed Precision Quantization

The inference stack is deeply co-designed with hardware, using Model-Optimizer to quantize routed-expert weights to NVFP4 (FP4), shared experts and Mamba projections to FP8, and caches to FP16/FP8 depending on context, supported by tailored calibration schemes (Four-Over-Six and MSE-based scaling). The result is full compatibility with Blackwell GPU FP4 hardware math and highly memory-efficient deployment.

Cache size and memory considerations are pivotal—at moderate sequence lengths, the Mamba SSM cache can outstrip classic FP8 KV cache unless aggressively quantized and optimized with periodic checkpointing (Figures 14, 15).

Figure 4: Cache size comparison for FP8 KV cache and Mamba SSM cache at different cache precisions.

Throughput and Serving Strategies

The hybrid Mamba-Attention + LatentMoE stack brings considerable inference benefits:

• Decode-heavy Regimes: Up to $1.6\times$ higher throughput than Qwen-3.5-397B-17B and $5.9\times$ over GLM-5.1-754B-A40B at 8K/64K decode workloads, while maintaining competitive accuracy (Figure 5).
• Speculative Decoding (MTP): MTP-based speculative decoding yields nearly $2.9\times$ throughput improvements at optimal draft length.
• Prefill-Decode Disaggregation: Supports optimal expert/data/tensor parallelism choices for prefill vs decode, maximizing both latency and throughput.
• Wide EP and TP: At large batch serving, EP dominates; for latency, wide TP is favorable.

  Figure 5: Accuracy and throughput comparisons—Nemotron 3 Ultra outperforms other open LLMs in throughput at high decode workloads without loss of accuracy.

---

Agentic, Reasoning, and Long-Context Evaluation

Comprehensive benchmarking demonstrates Nemotron 3 Ultra's strong agentic, reasoning, and long-context performance:

• Agentic Tasks: Terminal Bench 2.1, GDPVal, PinchBench, ProfBench, TAUBench, SWE-Bench all show Nemotron 3 Ultra competitive or near-leading scores.
• Reasoning/Math: High accuracy on IOI, LiveCodeBench, IMOAnswerBench, and GPQA (e.g., 570 on IOI 2025, 92.3 on IMOAnswerBench with tools). Robust test-time scaling on proof-centric math problems is observed, with accuracy exceeding 80% on Olympiad-level tasks.
• Long-Context: Maintains strong recall and reasoning up to 1M tokens on RULER, AA-LCR, and LongBench v2, facilitated by the sub-quadratic context scaling of the Mamba hybrid.
• Multilingual: Competitive on MMLU-ProX and WMT24++ for broad cross-lingual capability.

  Figure 6: Test-time scaling of Nemotron Nano 3 on Olympiad math, approaching the performance of math research agents.

---

Infrastructure and Scaling

The deployment of Nemotron 3 Ultra necessitated extensive infrastructure co-optimization:

• Ray GCS and Slurm Integration: Custom orchestration and node placement for efficient large-scale actor spawning and role assignment.
• NUMA and NVLink-Aware Placement: Precision placement to ensure expert parallelism and tensor parallelism maximize intra-node bandwidth.
• Optimized Caching and Initialization: JIT caches, container caching, and I/O minimization strategies reduced launch times at the scale of thousands of GPUs from nearly one hour to seconds.
• All-to-All Communication: Custom NVLinkOneSided and MoE chunking reduces MoE all-to-all bottlenecks.

---

Theoretical and Practical Implications

Nemotron 3 Ultra's release sets a new state-of-the-art in the public LLM space for agentic workloads, embedding design principles (hybrid Mamba-Attention, LatentMoE, MTP, and MOPD) that are likely to be foundational moving forward. The model's throughput scaling at sequence length and batch size makes it the preferred choice for long-running autonomous agents and ultra-long document reasoning. The rigorous analysis of training divergence, expert imbalance, and residual norm propagation at scale provides new insights into pretraining stability for sparse architectures.

The open-sourcing of all code, checkpoints (pre-trained, post-trained, quantized), and data cements Nemotron 3 Ultra as a platform for further research on scalable MoE architectures, inference acceleration, and agentic alignment.

---

Conclusion

Nemotron 3 Ultra demonstrates that highly-scaled, MoE hybrid LLMs can simultaneously deliver strong accuracy, agentic robustness, and unprecedented inference throughput by fundamentally co-designing architecture, data, training, quantization, and serving. This enables a new generation of open LLMs that are both efficient and reliable for the most demanding agentic workloads, setting a standard for future research and deployment in AI systems (2606.15007).