Saurav Muralidharan

Aaron Blakeman, Aarti Basant, Abhinav Khattar et al. · nvidia

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. We are releasing Nemotron-H base model checkpoints with support in Hugging Face and NeMo.

Aakshita Chandiramani, Aaron Blakeman, Abdullahi Olaoye et al. · amazon-science, cmu

We describe the pre-training, post-training, and quantization of Nemotron 3 Super, a 120 billion (active 12 billion) parameter hybrid Mamba-Attention Mixture-of-Experts model. Nemotron 3 Super is the first model in the Nemotron 3 family to 1) be pre-trained in NVFP4, 2) leverage LatentMoE, a new Mixture-of-Experts architecture that optimizes for both accuracy per FLOP and accuracy per parameter, and 3) include MTP layers for inference acceleration through native speculative decoding. We pre-trained Nemotron 3 Super on 25 trillion tokens followed by post-training using supervised fine tuning (SFT) and reinforcement learning (RL). The final model supports up to 1M context length and achieves comparable accuracy on common benchmarks, while also achieving up to 2.2x and 7.5x higher inference throughput compared to GPT-OSS-120B and Qwen3.5-122B, respectively. Nemotron 3 Super datasets, along with the base, post-trained, and quantized checkpoints, are open-sourced on HuggingFace.

Saurav Muralidharan, Sharath Turuvekere Sreenivas, Raviraj Joshi et al.

Large language models (LLMs) targeting different deployment scales and sizes are currently produced by training each variant from scratch; this is extremely compute-intensive. In this paper, we investigate if pruning an existing LLM and then re-training it with a fraction (<3%) of the original training data can be a suitable alternative to repeated, full retraining. To this end, we develop a set of practical and effective compression best practices for LLMs that combine depth, width, attention and MLP pruning with knowledge distillation-based retraining; we arrive at these best practices through a detailed empirical exploration of pruning strategies for each axis, methods to combine axes, distillation strategies, and search techniques for arriving at optimal compressed architectures. We use this guide to compress the Nemotron-4 family of LLMs by a factor of 2-4x, and compare their performance to similarly-sized models on a variety of language modeling tasks. Deriving 8B and 4B models from an already pretrained 15B model using our approach requires up to 40x fewer training tokens per model compared to training from scratch; this results in compute cost savings of 1.8x for training the full model family (15B, 8B, and 4B). Minitron models exhibit up to a 16% improvement in MMLU scores compared to training from scratch, perform comparably to other community models such as Mistral 7B, Gemma 7B and Llama-3 8B, and outperform state-of-the-art compression techniques from the literature. We have open-sourced Minitron model weights on Huggingface, with corresponding supplementary material including example code available on GitHub.

Sharath Turuvekere Sreenivas, Saurav Muralidharan, Raviraj Joshi et al.

We present a comprehensive report on compressing the Llama 3.1 8B and Mistral NeMo 12B models to 4B and 8B parameters, respectively, using pruning and distillation. We explore two distinct pruning strategies: (1) depth pruning and (2) joint hidden/attention/MLP (width) pruning, and evaluate the results on common benchmarks from the LM Evaluation Harness. The models are then aligned with NeMo Aligner and tested in instruct-tuned versions. This approach produces a compelling 4B model from Llama 3.1 8B and a state-of-the-art Mistral-NeMo-Minitron-8B (MN-Minitron-8B for brevity) model from Mistral NeMo 12B. We found that with no access to the original data, it is beneficial to slightly fine-tune teacher models on the distillation dataset. We open-source our base model weights on Hugging Face with a permissive license.

Gongfan Fang, Hongxu Yin, Saurav Muralidharan et al.

Large Language Models (LLMs) are distinguished by their massive parameter counts, which typically result in significant redundancy. This work introduces MaskLLM, a learnable pruning method that establishes Semi-structured (or ``N:M'') Sparsity in LLMs, aimed at reducing computational overhead during inference. Instead of developing a new importance criterion, MaskLLM explicitly models N:M patterns as a learnable distribution through Gumbel Softmax sampling. This approach facilitates end-to-end training on large-scale datasets and offers two notable advantages: 1) High-quality Masks - our method effectively scales to large datasets and learns accurate masks; 2) Transferability - the probabilistic modeling of mask distribution enables the transfer learning of sparsity across domains or tasks. We assessed MaskLLM using 2:4 sparsity on various LLMs, including LLaMA-2, Nemotron-4, and GPT-3, with sizes ranging from 843M to 15B parameters, and our empirical results show substantial improvements over state-of-the-art methods. For instance, leading approaches achieve a perplexity (PPL) of 10 or greater on Wikitext compared to the dense model's 5.12 PPL, but MaskLLM achieves a significantly lower 6.72 PPL solely by learning the masks with frozen weights. Furthermore, MaskLLM's learnable nature allows customized masks for lossless application of 2:4 sparsity to downstream tasks or domains. Code is available at https://github.com/NVlabs/MaskLLM.

Aarti Basant, Abhijit Khairnar, Abhijit Paithankar et al. · nvidia

We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

Aaron Blakeman, Aaron Grattafiori, Aarti Basant et al. · nvidia

We introduce the Nemotron 3 family of models - Nano, Super, and Ultra. These models deliver strong agentic, reasoning, and conversational capabilities. The Nemotron 3 family uses a Mixture-of-Experts hybrid Mamba-Transformer architecture to provide best-in-class throughput and context lengths of up to 1M tokens. Super and Ultra models are trained with NVFP4 and incorporate LatentMoE, a novel approach that improves model quality. The two larger models also include MTP layers for faster text generation. All Nemotron 3 models are post-trained using multi-environment reinforcement learning enabling reasoning, multi-step tool use, and support granular reasoning budget control. Nano, the smallest model, outperforms comparable models in accuracy while remaining extremely cost-efficient for inference. Super is optimized for collaborative agents and high-volume workloads such as IT ticket automation. Ultra, the largest model, provides state-of-the-art accuracy and reasoning performance. Nano is released together with its technical report and this white paper, while Super and Ultra will follow in the coming months. We will openly release the model weights, pre- and post-training software, recipes, and all data for which we hold redistribution rights.

Harvey Dam, Vinu Joseph, Aditya Bhaskara et al.

Network compression is now a mature sub-field of neural network research: over the last decade, significant progress has been made towards reducing the size of models and speeding up inference, while maintaining the classification accuracy. However, many works have observed that focusing on just the overall accuracy can be misguided. E.g., it has been shown that mismatches between the full and compressed models can be biased towards under-represented classes. This raises the important research question, can we achieve network compression while maintaining "semantic equivalence" with the original network? In this work, we study this question in the context of the "long tail" phenomenon in computer vision datasets observed by Feldman, et al. They argue that memorization of certain inputs (appropriately defined) is essential to achieving good generalization. As compression limits the capacity of a network (and hence also its ability to memorize), we study the question: are mismatches between the full and compressed models correlated with the memorized training data? We present positive evidence in this direction for image classification tasks, by considering different base architectures and compression schemes.

Salar Latifi, Saurav Muralidharan, Michael Garland

Transformer-based neural networks have achieved state-of-the-art task performance in a number of machine learning domains including natural language processing and computer vision. To further improve their accuracy, recent work has explored the integration of dynamic behavior into these networks in the form of mixture-of-expert (MoE) layers. In this paper, we explore the introduction of MoE layers to optimize a different metric: inference latency. We introduce a novel system named PLANER that takes an existing Transformer-based network and a user-defined latency target and produces an optimized, sparsely-activated version of the original network that tries to meet the latency target while maintaining baseline accuracy. We evaluate PLANER on two real-world language modeling tasks using the Transformer-XL network and achieve inference latency reductions of over 2x at iso-accuracy.

Cameron Shinn, Collin McCarthy, Saurav Muralidharan et al.

We introduce the Sparsity Roofline, a visual performance model for evaluating sparsity in neural networks. The Sparsity Roofline jointly models network accuracy, sparsity, and theoretical inference speedup. Our approach does not require implementing and benchmarking optimized kernels, and the theoretical speedup becomes equal to the actual speedup when the corresponding dense and sparse kernels are well-optimized. We achieve this through a novel analytical model for predicting sparse network performance, and validate the predicted speedup using several real-world computer vision architectures pruned across a range of sparsity patterns and degrees. We demonstrate the utility and ease-of-use of our model through two case studies: (1) we show how machine learning researchers can predict the performance of unimplemented or unoptimized block-structured sparsity patterns, and (2) we show how hardware designers can predict the performance implications of new sparsity patterns and sparse data formats in hardware. In both scenarios, the Sparsity Roofline helps performance experts identify sparsity regimes with the highest performance potential.

Sharath Turuvekere Sreenivas, Adithyakrishna Venkatesh Hanasoge, Mingyu Yang et al.

Cross-tokenizer knowledge distillation allows a student model to learn from teachers with incompatible vocabularies. Prior work operates on hidden states or logits; the latter is preferred as a drop-in replacement requiring no auxiliary components. Logit-based methods either use only the correct-token probability, missing the full 'dark knowledge' in the teacher's distribution, or operate on the full output distribution, relying on strict token partitioning and/or unprincipled heuristic ranking. We identify two key shortcomings of full-distribution, logit-based methods: (i) an uncommon-token failure, where critical tokens fall into the unmatched subset (e.g., Llama's 1100 multi-digit numerals under digit-splitting Qwen supervision) and are suppressed during training, reducing GSM8k from 12.89 to 2.56 compared to same-tokenizer KD from a weaker teacher; and (ii) over-conservative matching, where strict 1-to-1 matching excludes near-equivalent tokens across surface forms. These failures require distinct remedies: eliminating the partition when critical tokens are misaligned, and refining it when alignment is reliable. We propose X-Token, an approach with two complementary loss formulations targeting these issues. P-KL removes partitioning and aligns the student's distribution with the teacher's via a sparse projection matrix W (initialized from tokenizer-level string rules) to address the uncommon-token failure. H-KL retains the hybrid form while relaxing matching to align each student token with its top-ranked teacher mapping under W. Both objectives share W and extend naturally to multiple teachers. Empirically, on Llama-3.2-1B, X-Token outperforms the current state of the art GOLD by +3.82 average points with a Qwen3-4B teacher and by +0.5 with a Phi-4-Mini teacher. Further, a two-teacher setup (Phi-4-mini + Llama-3B) improves over single-teacher distillation by +1.3 points.

Shih-Yang Liu, Maksim Khadkevich, Nai Chit Fung et al.

While post-training compression techniques effectively reduce the memory footprint, latency, and power consumption of Large Language Models (LLMs), they often result in noticeable accuracy degradation and remain limited by hardware and kernel constraints that restrict supported compression formats ultimately reducing flexibility across a wide range of deployment scenarios. In this work, we propose EoRA, a novel fine-tuning-free method that augments compressed LLMs with low-rank matrices, allowing users to rapidly enhance task-specific performance and freely balance the trade-off between accuracy and computational overhead beyond the constraints of compression formats. EoRA consistently outperforms prior training-free low rank methods in recovering the accuracy of compressed LLMs, achieving notable accuracy improvements (e.g., $\mathbf{10.84\%}$ on ARC-Challenge, $\mathbf{6.74\%}$ on MathQA, and $\mathbf{6.74\%}$ on GSM8K) for LLaMA3-8B compressed to 3-bit. We also introduce an optimized CUDA kernel, accelerating inference by up to 1.4x and reducing memory overhead through quantizing EoRA. Overall, EoRA offers a prompt solution for improving the accuracy of compressed models under varying user requirements, enabling more efficient and flexible deployment of LLMs. Code is available at https://github.com/NVlabs/EoRA.

Vinu Joseph, Saurav Muralidharan, Animesh Garg et al.

Deep neural networks (DNNs) frequently contain far more weights, represented at a higher precision, than are required for the specific task which they are trained to perform. Consequently, they can often be compressed using techniques such as weight pruning and quantization that reduce both the model size and inference time without appreciable loss in accuracy. However, finding the best compression strategy and corresponding target sparsity for a given DNN, hardware platform, and optimization objective currently requires expensive, frequently manual, trial-and-error experimentation. In this paper, we introduce a programmable system for model compression called Condensa. Users programmatically compose simple operators, in Python, to build more complex and practically interesting compression strategies. Given a strategy and user-provided objective (such as minimization of running time), Condensa uses a novel Bayesian optimization-based algorithm to automatically infer desirable sparsities. Our experiments on four real-world DNNs demonstrate memory footprint and hardware runtime throughput improvements of 188x and 2.59x, respectively, using at most ten samples per search. We have released a reference implementation of Condensa at https://github.com/NVlabs/condensa.

Peter Belcak, Greg Heinrich, Shizhe Diao et al.

Large language models (LLMs) are often praised for exhibiting near-human performance on a wide range of tasks and valued for their ability to hold a general conversation. The rise of agentic AI systems is, however, ushering in a mass of applications in which language models perform a small number of specialized tasks repetitively and with little variation. Here we lay out the position that small language models (SLMs) are sufficiently powerful, inherently more suitable, and necessarily more economical for many invocations in agentic systems, and are therefore the future of agentic AI. Our argumentation is grounded in the current level of capabilities exhibited by SLMs, the common architectures of agentic systems, and the economy of LM deployment. We further argue that in situations where general-purpose conversational abilities are essential, heterogeneous agentic systems (i.e., agents invoking multiple different models) are the natural choice. We discuss the potential barriers for the adoption of SLMs in agentic systems and outline a general LLM-to-SLM agent conversion algorithm. Our position, formulated as a value statement, highlights the significance of the operational and economic impact even a partial shift from LLMs to SLMs is to have on the AI agent industry. We aim to stimulate the discussion on the effective use of AI resources and hope to advance the efforts to lower the costs of AI of the present day. Calling for both contributions to and critique of our position, we commit to publishing all such correspondence at https://research.nvidia.com/labs/lpr/slm-agents.

Ali Taghibakhshi, Ruisi Cai, Saurav Muralidharan et al.

Training a family of large language models (LLMs), either from scratch or via iterative compression, is prohibitively expensive and inefficient, requiring separate training runs for each model in the family. In this paper, we introduce Star Elastic, a novel LLM post-training method that adds N nested submodels to a given parent reasoning model using the compute of one run (N-fold savings) via a single post-training job. Beyond reducing training costs, Star Elastic also addresses a fundamental limitation of efficient reasoning: the rigidity of static architectures, which forces the allocation of constant resources regardless of token difficulty. By unlocking elastic budget control, Star Elastic enables a novel inference scheme that uses different submodels for each reasoning phase (thinking and answering). Star Elastic supports (1) nesting along the SSM, embedding channel, MoE, and FFN axes, (2) learning nested submodels via an end-to-end trainable router, and (3) curriculum-based knowledge distillation. Building on the Nemotron Elastic framework, we apply Star Elastic to the NVIDIA Nemotron Nano models, with a particular focus on hybrid Mixture-of-Experts (MoE) architectures: from Nemotron Nano v3 (30B/3.6A), we generate 23B (2.8A) and 12B (2.0A) variants with 160B training tokens. All nested models match or outperform independently trained baselines of comparable size and achieve a 360x reduction versus pretraining from scratch and a 7x reduction over state-of-the-art compression. Crucially, elastic budget control advances the accuracy-latency Pareto frontier, achieving up to 16% higher accuracy and 1.9x lower latency via dynamic per-phase model selection. We further extend Star Elastic to quantized regimes via Quantization-Aware Distillation (QAD), producing nested NVFP4 and FP8 elastic checkpoints that preserve zero-shot slicing while delivering smaller deployment footprints.

Ali Taghibakhshi, Sharath Turuvekere Sreenivas, Saurav Muralidharan et al.

Hybrid LLM architectures that combine Attention and State Space Models (SSMs) achieve state-of-the-art accuracy and runtime performance. Recent work has demonstrated that applying compression and distillation to Attention-only models yields smaller, more accurate models at a fraction of the training cost. In this work, we explore the effectiveness of compressing Hybrid architectures. We introduce a novel group-aware pruning strategy that preserves the structural integrity of SSM blocks and their sequence modeling capabilities. Furthermore, we demonstrate the necessity of such SSM pruning to achieve improved accuracy and inference speed compared to traditional approaches. Our compression recipe combines SSM, FFN, embedding dimension, and layer pruning, followed by knowledge distillation-based retraining, similar to the MINITRON technique. Using this approach, we compress the Nemotron-H 8B Hybrid model down to 4B parameters with up to 40x fewer training tokens. The resulting model surpasses the accuracy of similarly-sized models while achieving 2x faster inference, significantly advancing the Pareto frontier.

Ali Taghibakhshi, Sharath Turuvekere Sreenivas, Saurav Muralidharan et al.

Training a family of large language models targeting multiple scales and deployment objectives is prohibitively expensive, requiring separate training runs for each different size. Recent work on model compression through pruning and knowledge distillation has reduced this cost; however, this process still incurs hundreds of billions of tokens worth of training cost per compressed model. In this paper, we present Nemotron Elastic, a framework for building reasoning-oriented LLMs, including hybrid Mamba-Attention architectures, that embed multiple nested submodels within a single parent model, each optimized for different deployment configurations and budgets. Each of these submodels shares weights with the parent model and can be extracted zero-shot during deployment without additional training or fine-tuning. We enable this functionality through an end-to-end trained router, tightly coupled to a two-stage training curriculum designed specifically for reasoning models. We additionally introduce group-aware SSM elastification that preserves Mamba's structural constraints, heterogeneous MLP elastification, normalized MSE-based layer importance for improved depth selection, and knowledge distillation enabling simultaneous multi-budget optimization. We apply Nemotron Elastic to the Nemotron Nano V2 12B model, simultaneously producing a 9B and a 6B model using only 110B training tokens; this results in over 360x cost reduction compared to training model families from scratch, and around 7x compared to SoTA compression techniques. Each of the nested models performs on par or better than the SoTA in accuracy. Moreover, unlike other compression methods, the nested capability of our approach allows having a many-in-one reasoning model that has constant deployment memory against the number of models in the family.

Ruisi Cai, Saurav Muralidharan, Greg Heinrich et al.

Training modern LLMs is extremely resource intensive, and customizing them for various deployment scenarios characterized by limited compute and memory resources through repeated training is impractical. In this paper, we introduce Flextron, a network architecture and post-training model optimization framework supporting flexible model deployment. The Flextron architecture utilizes a nested elastic structure to rapidly adapt to specific user-defined latency and accuracy targets during inference with no additional fine-tuning required. It is also input-adaptive, and can automatically route tokens through its sub-networks for improved performance and efficiency. We present a sample-efficient training method and associated routing algorithms for systematically transforming an existing trained LLM into a Flextron model. We evaluate Flextron on the GPT-3 and LLama-2 family of LLMs, and demonstrate superior performance over multiple end-to-end trained variants and other state-of-the-art elastic networks, all with a single pretraining run that consumes a mere 7.63% tokens compared to original pretraining.

Yannan Nellie Wu, Po-An Tsai, Saurav Muralidharan et al.

Due to complex interactions among various deep neural network (DNN) optimization techniques, modern DNNs can have weights and activations that are dense or sparse with diverse sparsity degrees. To offer a good trade-off between accuracy and hardware performance, an ideal DNN accelerator should have high flexibility to efficiently translate DNN sparsity into reductions in energy and/or latency without incurring significant complexity overhead. This paper introduces hierarchical structured sparsity (HSS), with the key insight that we can systematically represent diverse sparsity degrees by having them hierarchically composed from multiple simple sparsity patterns. As a result, HSS simplifies the underlying hardware since it only needs to support simple sparsity patterns; this significantly reduces the sparsity acceleration overhead, which improves efficiency. Motivated by such opportunities, we propose a simultaneously efficient and flexible accelerator, named HighLight, to accelerate DNNs that have diverse sparsity degrees (including dense). Due to the flexibility of HSS, different HSS patterns can be introduced to DNNs to meet different applications' accuracy requirements. Compared to existing works, HighLight achieves a geomean of up to 6.4x better energy-delay product (EDP) across workloads with diverse sparsity degrees, and always sits on the EDP-accuracy Pareto frontier for representative DNNs

Vinu Joseph, Shoaib Ahmed Siddiqui, Aditya Bhaskara et al.

With the rise in edge-computing devices, there has been an increasing demand to deploy energy and resource-efficient models. A large body of research has been devoted to developing methods that can reduce the size of the model considerably without affecting the standard metrics such as top-1 accuracy. However, these pruning approaches tend to result in a significant mismatch in other metrics such as fairness across classes and explainability. To combat such misalignment, we propose a novel multi-part loss function inspired by the knowledge-distillation literature. Through extensive experiments, we demonstrate the effectiveness of our approach across different compression algorithms, architectures, tasks as well as datasets. In particular, we obtain up to $4.1\times$ reduction in the number of prediction mismatches between the compressed and reference models, and up to $5.7\times$ in cases where the reference model makes the correct prediction; all while making no changes to the compression algorithm, and minor modifications to the loss function. Furthermore, we demonstrate how inducing simple alignment between the predictions of the models naturally improves the alignment on other metrics including fairness and attributions. Our framework can thus serve as a simple plug-and-play component for compression algorithms in the future.