We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use. The family comes in three sizes -- Nano (8B), Super (49B), and Ultra (253B) -- and performs competitively with state-of-the-art reasoning models such as DeepSeek-R1 while offering superior inference throughput and memory efficiency. In this report, we discuss the training procedure for these models, which entails using neural architecture search from Llama 3 models for accelerated inference, knowledge distillation, and continued pretraining, followed by a reasoning-focused post-training stage consisting of two main parts: supervised fine-tuning and large scale reinforcement learning. Llama-Nemotron models are the first open-source models to support a dynamic reasoning toggle, allowing users to switch between standard chat and reasoning modes during inference. To further support open research and facilitate model development, we provide the following resources: 1. We release the Llama-Nemotron reasoning models -- LN-Nano, LN-Super, and LN-Ultra -- under the commercially permissive NVIDIA Open Model License Agreement. 2. We release the complete post-training dataset: Llama-Nemotron-Post-Training-Dataset. 3. We also release our training codebases: NeMo, NeMo-Aligner, and Megatron-LM.
核心贡献 · Key contributions
提出 Llama-Nemotron，一个开放的非同构推理模型系列，包含三种规模（8B、49B、253B）。 Introduces Llama-Nemotron, an open family of heterogeneous reasoning models with three sizes (8B, 49B, 253B).
在推理性能上与最先进模型竞争，同时提供更优的推理吞吐量和内存效率。 Achieves competitive reasoning performance with state-of-the-art models while offering superior inference throughput and memory efficiency.
首个支持动态推理切换的开源模型，可在标准聊天和推理模式间切换。 First open-source models to support a dynamic reasoning toggle for switching between standard chat and reasoning modes.
在宽松许可下发布完整的后训练数据集、训练代码库（NeMo、NeMo-Aligner、Megatron-LM）和模型权重。 Releases complete post-training dataset, training codebases (NeMo, NeMo-Aligner, Megatron-LM), and model weights under permissive license.
证明大规模课程驱动的强化学习使学生模型能够超越教师模型的性能。 Demonstrates that large-scale curriculum-driven reinforcement learning enables student models to surpass teacher performance.
在 LLM 作为评判者等分布外任务上表现出强泛化能力，超越许多专有和开源模型。 Shows strong generalization on out-of-distribution tasks like LLM-as-a-Judge, outperforming many proprietary and open-source models.
局限 · Limitations
对于较小模型，强化学习相比蒸馏效果欠佳。 Reinforcement learning yields suboptimal results for smaller models compared to distillation.
LN-Ultra 的强化学习训练中出现不稳定性（梯度爆炸），需重新初始化优化器状态后恢复训练。 Training instability (gradient explosions) observed during RL for LN-Ultra, requiring resumption with reinitialized optimizer states.
在 AIME 等推理密集型任务上的评估因数据集小和生成随机性而呈现高方差。 Evaluation on reasoning-heavy tasks like AIME exhibits high variance due to small dataset size and generation randomness.
仅靠 SFT 会限制学生模型性能上限；超越教师需要 RL，但资源消耗大。 SFT alone sets an upper bound on student performance; RL is necessary to surpass teacher but resource-intensive.
模型针对特定硬件（H100 GPU）优化；在其他硬件上的性能可能有所不同。 Models are optimized for specific hardware (H100 GPUs); performance on other hardware may vary.
论文章节 · Sections（共 21）
摘要Abstract
1 引言1 Introduction
2 创建推理优化模型2 Creating Inference-Optimized Models
2.1 部署约束与效率目标2.1 Deployment Constraints and Efficiency Targets
2.2 后 NAS 训练：知识蒸馏与持续预训练2.2 Post-NAS Training: Knowledge Distillation and Continued Pretraining