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

• URL: http://arxiv.org/abs/2511.18890v1
• Date: Mon, 24 Nov 2025 08:46:36 GMT
• Title: Nemotron-Flash: Towards Latency-Optimal Hybrid Small Language Models
• Authors: 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,
• Abstract summary: This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training.<n>We introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy-efficiency frontier of state-of-the-art SLMs.
• Score: 97.55009021098554
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Efficient deployment of small language models (SLMs) is essential for numerous real-world applications with stringent latency constraints. While previous work on SLM design has primarily focused on reducing the number of parameters to achieve parameter-optimal SLMs, parameter efficiency does not necessarily translate into proportional real-device speed-ups. This work aims to identify the key determinants of SLMs' real-device latency and offer generalizable principles and methodologies for SLM design and training when real-device latency is the primary consideration. Specifically, we identify two central architectural factors: depth-width ratios and operator choices. The former is crucial for small-batch-size latency, while the latter affects both latency and large-batch-size throughput. In light of this, we first study latency-optimal depth-width ratios, with the key finding that although deep-thin models generally achieve better accuracy under the same parameter budget, they may not lie on the accuracy-latency trade-off frontier. Next, we explore emerging efficient attention alternatives to evaluate their potential as candidate building operators. Using the identified promising operators, we construct an evolutionary search framework to automatically discover latency-optimal combinations of these operators within hybrid SLMs, thereby advancing the accuracy-latency frontier. In addition to architectural improvements, we further enhance SLM training using a weight normalization technique that enables more effective weight updates and improves final convergence. Combining these methods, we introduce a new family of hybrid SLMs, called Nemotron-Flash, which significantly advances the accuracy-efficiency frontier of state-of-the-art SLMs, e.g., achieving over +5.5% average accuracy, 1.3x/1.9x lower latency, and 18.7x/45.6x higher throughput compared to Qwen3-1.7B/0.6B, respectively.

Related papers

• Optimal Batch-Size Control for Low-Latency Federated Learning with Device Heterogeneity [30.53961003855146]
  Federated learning (FL) has emerged as a popular approach for collaborative machine learning in sixth-generation (6G) networks.<n>The deployment of FL algorithms is expected to empower a wide range of Internet-of-Things (IoT) applications, e.g., autonomous driving, augmented reality, and healthcare.<n>We propose a novel C$2$-aware framework for optimal batch-size control that minimizes end-to-end (E2E) learning latency while ensuring convergence.
  arXiv  Detail & Related papers  (2025-07-21T13:24:38Z)
• Optuna vs Code Llama: Are LLMs a New Paradigm for Hyperparameter Tuning? [45.58422897857411]
  This work explores the use of large language models (LLMs) for hyperparameter optimization by fine-tuning a parameter-efficient version of Code Llama using LoRA.<n>Our approach achieves competitive or superior Root Mean Square Error (RMSE) while substantially reducing computational overhead.<n>Results demonstrate that LLM-based optimization not only rivals established Bayesian methods like Tree-structured Parzen Estimators (TPE) but also accelerates tuning for real-world applications requiring perceptual quality and low-latency processing.
  arXiv  Detail & Related papers  (2025-04-08T13:15:47Z)
• MOFHEI: Model Optimizing Framework for Fast and Efficient Homomorphically Encrypted Neural Network Inference [0.8388591755871735]
  Homomorphic Encryption (HE) enables us to perform machine learning tasks over encrypted data.<n>We propose MOFHEI, a framework that optimize the model to make HE-based neural network inference, fast and efficient.<n>Our framework achieves up to 98% pruning ratio on LeNet, eliminating up to 93% of the required HE operations for performing PI.
  arXiv  Detail & Related papers  (2024-12-10T22:44:54Z)
• Read-ME: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design [59.00758127310582]
  We propose a novel framework Read-ME that transforms pre-trained dense LLMs into smaller MoE models. Our approach employs activation sparsity to extract experts. Read-ME outperforms other popular open-source dense models of similar scales.
  arXiv  Detail & Related papers  (2024-10-24T19:48:51Z)
• ShiftAddLLM: Accelerating Pretrained LLMs via Post-Training Multiplication-Less Reparameterization [13.622268474310918]
  ShiftAddLLM is an efficient multiplication-free model for large language models. It achieves perplexity improvements of 5.6 and 22.7 points at comparable or lower latency. Experiments on five LLM families and eight tasks consistently validate the effectiveness of ShiftAddLLM.
  arXiv  Detail & Related papers  (2024-06-10T02:47:55Z)
• HOPE for a Robust Parameterization of Long-memory State Space Models [51.66430224089725]
  State-space models (SSMs) that utilize linear, time-invariant (LTI) systems are known for their effectiveness in learning long sequences. We develop a new parameterization scheme, called HOPE, for LTI systems that utilize Markov parameters within Hankel operators. Our new parameterization endows the SSM with non-decaying memory within a fixed time window, which is empirically corroborated by a sequential CIFAR-10 task with padded noise.
  arXiv  Detail & Related papers  (2024-05-22T20:20:14Z)
• Not All Attention is Needed: Parameter and Computation Efficient Transfer Learning for Multi-modal Large Language Models [73.48675708831328]
  We propose a novel parameter and computation efficient tuning method for Multi-modal Large Language Models (MLLMs) The Efficient Attention Skipping (EAS) method evaluates the attention redundancy and skips the less important MHAs to speed up inference. The experiments show that EAS not only retains high performance and parameter efficiency, but also greatly speeds up inference speed.
  arXiv  Detail & Related papers  (2024-03-22T14:20:34Z)
• MobiLlama: Towards Accurate and Lightweight Fully Transparent GPT [87.4910758026772]
  "Bigger the better" has been the predominant trend in recent Large Language Models (LLMs) development. This paper explores the "less is more" paradigm by addressing the challenge of designing accurate yet efficient Small Language Models (SLMs) for resource constrained devices.
  arXiv  Detail & Related papers  (2024-02-26T18:59:03Z)
• Collaborative Intelligent Reflecting Surface Networks with Multi-Agent Reinforcement Learning [63.83425382922157]
  Intelligent reflecting surface (IRS) is envisioned to be widely applied in future wireless networks. In this paper, we investigate a multi-user communication system assisted by cooperative IRS devices with the capability of energy harvesting.
  arXiv  Detail & Related papers  (2022-03-26T20:37:14Z)
• Particle Swarm Optimized Federated Learning For Industrial IoT and Smart City Services [9.693848515371268]
  We propose a Particle Swarm Optimization (PSO)-based technique to optimize the hyper parameter settings for the local Machine Learning models. We evaluate the performance of our proposed technique using two case studies.
  arXiv  Detail & Related papers  (2020-09-05T16:20:47Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.