Llama-Nemotron: Efficient Reasoning Models

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.

Nemotron-H: A Family of Accurate and Efficient Hybrid Mamba-Transformer Models

As inference-time scaling becomes critical for enhanced reasoning capabilities, it is increasingly becoming important to build models that are efficient to infer. We introduce Nemotron-H, a family of 8B and 56B/47B hybrid Mamba-Transformer models designed to reduce inference cost for a given accuracy level. To achieve this goal, we replace the majority of self-attention layers in the common Transformer model architecture with Mamba layers that perform constant computation and require constant memory per generated token. We show that Nemotron-H models offer either better or on-par accuracy compared to other similarly-sized state-of-the-art open-sourced Transformer models (e.g., Qwen-2.5-7B/72B and Llama-3.1-8B/70B), while being up to 3$\times$ faster at inference. To further increase inference speed and reduce the memory required at inference time, we created Nemotron-H-47B-Base from the 56B model using a new compression via pruning and distillation technique called MiniPuzzle. Nemotron-H-47B-Base achieves similar accuracy to the 56B model, but is 20% faster to infer. In addition, we introduce an FP8-based training recipe and show that it can achieve on par results with BF16-based training. This recipe is used to train the 56B model. All Nemotron-H models will be released, with support in Hugging Face, NeMo, and Megatron-LM.

STANCE-C3: Domain-adaptive Cross-target Stance Detection via Contrastive Learning and Counterfactual Generation

Stance detection is the process of inferring a person's position or standpoint on a specific issue to deduce prevailing perceptions toward topics of general or controversial interest, such as health policies during the COVID-19 pandemic. Existing models for stance detection are trained to perform well for a single domain (e.g., COVID-19) and a specific target topic (e.g., masking protocols), but are generally ineffectual in other domains or targets due to distributional shifts in the data. However, constructing high-performing, domain-specific stance detection models requires an extensive corpus of labeled data relevant to the targeted domain, yet such datasets are not readily available. This poses a challenge as the process of annotating data is costly and time-consuming. To address these challenges, we introduce a novel stance detection model coined domain-adaptive Cross-target STANCE detection via Contrastive learning and Counterfactual generation (STANCE-C3) that uses counterfactual data augmentation to enhance domain-adaptive training by enriching the target domain dataset during the training process and requiring significantly less information from the new domain. We also propose a modified self-supervised contrastive learning as a component of STANCE-C3 to prevent overfitting for the existing domain and target and enable cross-target stance detection. Through experiments on various datasets, we show that STANCE-C3 shows performance improvement over existing state-of-the-art methods.