Making intelligence
efficient.

EdgeWise: Bringing Long-Horizon Agents to the Edge via Selective Cloud Delegation

Zeru Shi, Jindong Jiang, Ryo Hachiuma, Yonggan Fu, Byung-Kwan Lee, Peter Belcak, Saurav Muralidharan, Ayush Agarwal, Ruixiang Tang, Ambrish Dantrey, Pavlo Molchanov, Anjul Patney, Subhashree Radhakrishnan

EdgeWise runs a small language model locally as the primary agent and delegates critical steps to a cloud model. Step-level reinforcement learning teaches the local model when to delegate, while an external judge offers a training-free alternative. Across long-horizon coding tasks, EdgeWise matches or exceeds cloud-only performance while reducing inference costs by up to 62% on SWE-Rebench and 34% on Terminal-Bench 2.1.

Arxiv ↗Project Page ↗

EMNLP 2026

Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding

Yonggan Fu, Lexington Whalen, Abhinav Garg, Chengyue Wu, Maksim Khadkevich, Nicolai Oswald, Enze Xie, Daniel Egert, Sharath Turuvekere Sreenivas, Shizhe Diao, Chenhan Yu, Ye Yu, Weijia Chen, Sajad Norouzi, Jingyu Liu, Shiyi Lan, Ligeng Zhu, Jin Wang, Jindong Jiang, Morteza Mardani, Mehran Maghoumi, Song Han, Ante Jukić, Nima Tajbakhsh, Jan Kautz, Pavlo Molchanov

Nemotron-Labs-Diffusion unifies autoregressive, diffusion, and self-speculation decoding within a single language model trained with a joint AR-diffusion objective. The resulting 3B, 8B, and 14B model family can switch decoding modes across deployment settings, delivering strong efficiency across different serving concurrency levels and outperforming existing multi-token prediction methods. By combining the accuracy of autoregressive generation with diffusion-based parallelism, Nemotron-Labs-Diffusion-8B achieves 4x higher throughput than Qwen3-8B on SPEED-Bench using SGLang on an NVIDIA GB200 GPU.

Arxiv ↗Models ↗Project Page ↗

ICML 2026

RADIO1D: Elastic Representations for Condensed Vision Modeling

Greg Heinrich, Mike Ranzinger, Collin McCarthy, Natan Bagrov, Eugene Khvedchenya, Bryan Catanzaro, Jan Kautz, Andrew Tao, Pavlo Molchanov

RADIO1D compresses images into a compact, variable-length 1D token sequence, enabling accurate scene understanding with as little as one token and flexible accuracy-efficiency tradeoffs for vision-language models.

Arxiv ↗Code ↗

Nemotron-Labs-Diffusion-Image: Advancing Masked Discrete Diffusion for High-Resolution Image Synthesis

Shufan Li, Greg Heinrich, Hanrong Ye, Yonggan Fu, Aditya Grover, Jan Kautz, Pavlo Molchanov

NL-Diffusion-Image generates high-resolution images through masked discrete diffusion, using grouped cross-entropy and token editing to improve training efficiency, image fidelity, and self-correction while supporting flexible few-step generation.

Arxiv ↗Project Page ↗Models ↗Demo ↗

ICML 2026

Star Elastic: Many-in-One Reasoning LLMs with Efficient Budget Control

Ali Taghibakhshi, Ruisi Cai, Saurav Muralidharan, Sharath Turuvekere Sreenivas, Aditya Vavre, Ameya Sunil Mahabaleshwarkar, Bilal Kartal, Sheldon Liang, Marcin Chochowski, Zijia Chen, Akhiad Bercovich, Ran Zilberstein, Ran El-Yaniv, Yonatan Geifman, Daniel Korzekwa, Yoshi Suhara, Oluwatobi Olabiyi, Ashwath Aithal, Nima Tajbakhsh, Pavlo Molchanov

Star Elastic is a post-training method that takes an already-trained reasoning LLM and, in a single training job, embeds N nested submodels inside it that share its weights and can be sliced out at deployment with no further training; this yields N-fold compute savings over training each model separately. It also unlocks elastic budget control: because the submodels are nested, inference can use a different submodel for each reasoning phase, rather than allocating constant resources regardless of token difficulty. Applied to the pretrained Nemotron 3 Nano 30B-A3B, it produced a 30B + 23B + 12B nested family with only ~160B tokens; the Elastic-30B variant matches or exceeds the parent on most benchmarks, and hybrid budget control delivers up to 16% higher accuracy and 1.9x faster inference than standard single-model budget control.

Arxiv ↗Models ↗

ICML 2026

GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization

Shih-Yang Liu, Xin Dong, Ximing Lu, Shizhe Diao, Peter Belcak, Mingjie Liu, Min-Hung Chen, Hongxu Yin, Yu-Chiang Frank Wang, Kwang-Ting Cheng, Yejin Choi, Jan Kautz, Pavlo Molchanov

GDPO improves GRPO for reinforcement learning with multiple rewards by normalizing each reward separately before combining them. This preserves distinctions that GRPO can collapse when it sums rewards first, while final batch normalization supports stable optimization across correctness, formatting, length, and other objectives.

Arxiv ↗Project Page ↗Code ↗Hugging Face ↗

ICML 2026

Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed

Yonggan Fu, Lexington Whalen, Zhifan Ye, Xin Dong, Shizhe Diao, Jingyu Liu, Chengyue Wu, Hao Zhang, Enze Xie, Song Han, Maksim Khadkevich, Jan Kautz, Yingyan Celine Lin, Pavlo Molchanov

Efficient-DLM converts pretrained autoregressive language models into diffusion language models that preserve strong task accuracy while enabling parallel generation. It introduces a continuous pretraining scheme with block-wise attention that remains causal across blocks while enabling bidirectional modeling within each block, together with position-dependent token masking to better align training with diffusion decoding. The resulting 4B and 8B model family advances both accuracy and efficiency, with Efficient-DLM-8B achieving +5.4%/+2.7% higher accuracy and 4.5x/2.7x higher throughput than Dream-7B and Qwen3-4B, respectively.

Arxiv ↗Models ↗

NeurIPS 2025

Nemotron-Flash: Towards Latency-Optimal Hybrid Small Language Models

Yonggan Fu, Xin Dong, Shizhe Diao, Matthijs Van Keirsbilck, Hanrong Ye, Wonmin Byeon, Yashaswi Karnati, Lucas Liebenwein, Hannah Zhang, Nikolaus Binder, Maksim Khadkevich, Alexander Keller, Jan Kautz, Yingyan Celine Lin, Pavlo Molchanov

Nemotron-Flash rethinks small language model design around real-device latency rather than parameter count. It identifies depth-width ratios and operator choices as key determinants of inference efficiency, combining latency-optimal depth-width design, evolutionary search over hybrid operators, and training-time weight normalization. The resulting 1B and 3B model family advances the accuracy-efficiency frontier across both low-latency and high-throughput serving.

Arxiv ↗Models ↗

ICLR 2026

OmniVinci: Enhancing Architecture and Data for Omni-Modal Understanding LLM

Hanrong Ye, ..., Jan Kautz, Hongxu Yin, Pavlo Molchanov

OmniVinci is NVIDIA's first systematic research of new model architecture and data curation for omni-modal LLMs, resulting in a model that achieves state-of-the-art performance in joint perception of images, videos, audio, and text.

Arxiv ↗Project Page ↗Code ↗Models ↗

Minifinetuning: Low-Data Generation Domain Adaptation through Corrective Self-Distillation

Peter Belcak, Greg Heinrich, Jan Kautz, Pavlo Molchanov

Finetuning language models for a new domain inevitably leads to the deterioration of their general performance. This becomes more pronounced the more limited the finetuning data resource.We introduce minifinetuning (MFT), a method for language model domain adaptation that considerably reduces the effects of overfitting-induced degeneralization in low-data settings and which does so in the absence of any pre-training data for replay. MFT demonstrates 2-10x more favourable specialization-to-degeneralization ratios than standard finetuning across a wide range of models and domains and exhibits an intrinsic robustness to overfitting when data in the new domain is scarce and down to as little as 500 samples.Employing corrective self-distillation that is individualized on the sample level, MFT outperforms parameter-efficient finetuning methods, demonstrates replay-like degeneralization mitigation properties, and is composable with either for a combined effect.

Arxiv ↗Project Page ↗

Small Language Models are the Future of Agentic AI

Peter Belcak, Greg Heinrich, Shizhe Diao, Yonggan Fu, Xin Dong, Saurav Muralidharan, Yingyan Celine Lin, Pavlo Molchanov

Small language models (SLMs) are often overlooked in favor of larger, more powerful models. However, we argue that small LLMs are the future of agentic AI due to their efficiency, accessibility, and adaptability. Here we detail our position on the future of agentic AI, arguing with evidence that that the industry and the community are to benefit from a switch from LLMs to SLMs. Our position is based on three key points: That SLMs are sufficiently powerful, inherently more suitable, and necessarily more economical for many invocations in agentic systems. We then draw a conclusion that they are, to a large extent, the future of agentic AI. We also outline a LLM-to-SLM conversion algorithm that can be employed to produce a SLM for a specific agentic workflow. We welcome correspondence on the topic.

Arxiv ↗Project/Correspondence Page ↗

NeurIPS 2025 (spotlight)

CLIMB: CLustering-based Iterative Data Mixture Bootstrapping

Shizhe Diao, Yu Yang, Yonggan Fu, Xin Dong, Dan Su, Markus Kliegl, Zijia Chen, Peter Belcak, Yoshi Suhara, Hongxu Yin, Mostofa Patwary, Yingyan (Celine) Lin, Jan Kautz, Pavlo Molchanov

We introduce CLustering-based Iterative Data Mixture Bootstrapping (CLIMB), an automated framework that discovers, evaluates, and refines data mixtures in a pre-training setting. Specifically, CLIMB embeds and clusters large-scale datasets in a semantic space and then iteratively searches for optimal mixtures using a smaller proxy model and a predictor. When continuously trained on 400B tokens with this mixture, our 950M model exceeds the state-of-the-art Llama-3.2-1B by 2.0% averaged across 12 general reasoning tasks. Moreover, we observe that optimizing for a specific domain (e.g., Social Sciences) yields a 5% improvement over random sampling. We introduce ClimbLab, a filtered 1.3-trillion-token corpus with 20 clusters as a research playground, and ClimbMix, a compact yet powerful 400-billion-token dataset designed for efficient pre-training that delivers superior performance under an equal token budget.

Arxiv ↗HF datasets ↗Homepage ↗

CVPR 2025 (highlight)

VILA-HD: Scaling Vision Pre-Training to 4K Resolution

Baifeng Shi, Boyi Li, Han Cai, Yao Lu, Sifei Liu, Marco Pavone, Jan Kautz, Song Han, Trevor Darrell, Pavlo Molchanov, Hongxu Yin

We introduce PS3, a framework that scales CLIP-style vision pre-training to 4K resolution with near-constant cost by selectively processing local regions and contrasting them with detailed captions. This enables high-resolution representation learning with greatly reduced computational overhead. The resulting model, VILA-HD, significantly improves high-resolution visual perception compared to baselines without high-resolution vision pre-training, such as AnyRes and S², while using up to 4.3× fewer tokens. VILA-HD outperforms previous multi-modal large language models (MLLMs), including NVILA and Qwen2-VL, across multiple benchmarks. To further evaluate high-resolution perception, we propose 4KPro, a new benchmark of image QA at 4K resolution, on which VILA-HD achieves a 14.5% improvement over GPT-4o and a 3.2% improvement with 2.96× speedup over Qwen2-VL.

Arxiv ↗PubPage ↗

ICML 2025

FeatSharp: Your Vision Model Features, Sharper

Mike Ranzinger, Greg Heinrich, Pavlo Molchanov, Jan Kautz, Bryan Catanzaro, Andrew Tao

The feature maps of vision encoders are fundamental to myriad modern AI tasks, ranging from core perception algorithms (e.g. semantic segmentation, object detection, depth perception, etc.) to modern multimodal understanding in vision-language models (VLMs). Currently, in computer vision, the frontier of general purpose vision backbones are Vision Transformers (ViT), typically trained using contrastive loss (e.g. CLIP). A key problem with most off-the-shelf ViTs, particularly CLIP, is that these models are inflexibly low resolution. Most run at 224x224px, while the "high resolution" versions are around 378-448px, but still inflexible. We introduce a novel method to coherently and cheaply upsample the feature maps of low-res vision encoders while picking up on fine-grained details that would otherwise be lost due to resolution. We demonstrate the effectiveness of this approach on core perception tasks as well as within agglomerative model (RADIO) training as a way of providing richer targets for distillation.

Arxiv ↗

CVPR 2025

RADIOv2.5: Improved Baselines for Agglomerative Vision Foundation Models

Greg Heinrich, Mike Ranzinger, Hongxu (Danny)Yin, Yao Lu, Jan Kautz, Andrew Tao, Bryan Catanzaro, Pavlo Molchanov

Agglomerative models have recently emerged as a powerful approach to training vision foundation models, leveraging multi-teacher distillation from existing models such as CLIP, DINO, and SAM. This strategy enables the efficient creation of robust models, combining the strengths of individual teachers while significantly reducing computational and resource demands. In this paper, we thoroughly analyze state-of-the-art agglomerative models, identifying critical challenges including resolution mode shifts, teacher imbalance, idiosyncratic teacher artifacts, and an excessive number of output tokens. To address these issues, we propose several novel solutions: multi-resolution training, mosaic augmentation, and improved balancing of teacher loss functions. Specifically, in the context of Vision Language Models, we introduce a token compression technique to maintain high-resolution information within a fixed token count.

Arxiv ↗

CVPR 2025

NVILA: Efficient Frontier Visual Language Models

Zhijian Liu, Ligeng Zhu, Baifeng Shi, Zhuoyang Zhang, Yuming Lou, Shang Yang, Haocheng Xi, Shiyi Cao, Yuxian Gu, Dacheng Li, Xiuyu Li, Yunhao Fang, Yukang Chen, Cheng-Yu Hsieh, De-An Huang, An-Chieh Cheng, Vishwesh Nath, Jinyi Hu, Sifei Liu, Ranjay Krishna, Daguang Xu, Xiaolong Wang, Pavlo Molchanov, Jan Kautz, Hongxu Yin, Song Han, Yao Lu

We introduce NVILA, a family of open visual language models (VLMs) designed to optimize both efficiency and accuracy. Building upon VILA, NVILA employs a 'scale-then-compress' strategy: first enhancing spatial and temporal resolutions to preserve visual details, then compressing visual tokens to reduce computational overhead. This approach enables NVILA to efficiently process high-resolution images and long videos. Our systematic investigation enhances NVILA's efficiency throughout its lifecycle, from training and fine-tuning to deployment. NVILA matches or surpasses the accuracy of many leading open and proprietary VLMs across a wide range of image and video benchmarks. Simultaneously, it reduces training costs by 4.5×, fine-tuning memory usage by 3.4×, pre-filling latency by 1.6–2.2×, and decoding latency by 1.2–2.8×.

Arxiv ↗Project Page ↗GitHub ↗

ICLR 2025 (spotlight)

Hymba: A Hybrid-head Architecture for Small Language Models

Xin Dong*, Yonggan Fu*, Shizhe Diao, Wonmin Byeon, Zijia Chen, Ameya Sunil Mahabaleshwarkar, Shih-Yang Liu, Matthijs Van Keirsbilck, Min-Hung Chen, Yoshi Suhara, Yingyan Lin, Jan Kautz, Pavlo Molchanov

We propose Hymba, a family of small language models featuring a hybrid-head parallel architecture that integrates transformer attention mechanisms with state space models (SSMs) for enhanced efficiency. Attention heads provide high-resolution recall, while SSM heads enable efficient context summarization. Additionally, we introduce learnable meta tokens that are prepended to prompts, storing critical information and alleviating the "forced-to-attend" burden associated with attention mechanisms. This model is further optimized by incorporating cross-layer key-value (KV) sharing and partial sliding window attention, resulting in a compact cache size. Notably, Hymba achieves state-of-the-art results for small LMs.

Arxiv ↗HF models ↗Blog ↗

NeurIPS 2024 (spotlight)

MaskLLM: Learnable Semi-Structured Sparsity for Large Language Models

Gongfan Fang, Hongxu Yin, Saurav Muralidharan, Greg Heinrich, Jeff Pool, Jan Kautz, Pavlo Molchanov, Xinchao Wang

MaskLLM learns 2:4 semi-structured sparsity end to end by modeling pruning masks as differentiable distributions while keeping model weights frozen. Across LLaMA-2, Nemotron-4, and GPT-3 models from 843M to 15B parameters, it preserves substantially more accuracy than prior pruning methods and supports transferable, task-specific masks for efficient inference.

Arxiv ↗Code ↗

NeurIPS 2024

Compact Language Models via Pruning and Knowledge Distillation

Saurav Muralidharan, Sharath Turuvekere Sreenivas, Raviraj Joshi, Marcin Chochowski, Mostofa Patwary, Mohammad Shoeybi, Bryan Catanzaro, Jan Kautz, Pavlo Molchanov

We develop an efficient model compression strategy for LLMs that combines depth, width, attention and MLP pruning with knowledge-distillation-based retraining. We use our strategy 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).

Arxiv ↗HF models ↗Blog ↗

ICML 2024 (oral)

Flextron: Many-in-One Flexible Large Language Model

Ruisi Cai, Saurav Muralidharan, Greg Heinrich, Hongxu Yin, Zhangyang Wang, Jan Kautz, Pavlo Molchanov

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. We introduce Flextron, a network architecture and post-training model optimization framework supporting flexible model deployment.

Arxiv ↗Webpage ↗

ICML 2024 (oral)

DoRA: Weight-Decomposed Low-Rank Adaptation

Shih-Yang Liu, Chien-Yi Wang, Hongxu Yin, Pavlo Molchanov, Yu-Chiang Frank Wang, Kwang-Ting Cheng, Min-Hung Chen

DoRA improves low-rank adaptation by decomposing each pretrained weight into magnitude and direction, training the magnitude directly while using LoRA for directional updates. Across LLaMA, LLaVA, and VL-BART, it more closely resembles full fine-tuning and consistently outperforms LoRA while retaining low parameter costs and adding no inference latency.

Arxiv ↗Project Page ↗Code ↗Blog ↗