MobileLLM-Flash: Latency-Guided On-Device LLM Design for Industry Scale Deployment

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Huang, Hanxian, Fedorov, Igor, Gromov, Andrey, Beckerman, Bernard, Suda, Naveen, Eriksson, David, Balandat, Maximilian, Conway, Rylan, Huber, Patrick, Sankar, Chinnadhurai, Dalmia, Ayushi, Liu, Zechun, Wu, Lemeng, Elgamal, Tarek, Sagar, Adithya, Chandra, Vikas, Krishnamoorthi, Raghuraman
Format: Preprint
Veröffentlicht: 2026
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866917441818329088
author Huang, Hanxian
Fedorov, Igor
Gromov, Andrey
Beckerman, Bernard
Suda, Naveen
Eriksson, David
Balandat, Maximilian
Conway, Rylan
Huber, Patrick
Sankar, Chinnadhurai
Dalmia, Ayushi
Liu, Zechun
Wu, Lemeng
Elgamal, Tarek
Sagar, Adithya
Chandra, Vikas
Krishnamoorthi, Raghuraman
author_facet Huang, Hanxian
Fedorov, Igor
Gromov, Andrey
Beckerman, Bernard
Suda, Naveen
Eriksson, David
Balandat, Maximilian
Conway, Rylan
Huber, Patrick
Sankar, Chinnadhurai
Dalmia, Ayushi
Liu, Zechun
Wu, Lemeng
Elgamal, Tarek
Sagar, Adithya
Chandra, Vikas
Krishnamoorthi, Raghuraman
contents Real-time AI experiences call for on-device large language models (OD-LLMs) optimized for efficient deployment on resource-constrained hardware. The most useful OD-LLMs produce near-real-time responses and exhibit broad hardware compatibility, maximizing user reach. We present a methodology for designing such models using hardware-in-the-loop architecture search under mobile latency constraints. This system is amenable to industry-scale deployment: it generates models deployable without custom kernels and compatible with standard mobile runtimes like Executorch. Our methodology avoids specialized attention mechanisms and instead uses attention skipping for long-context acceleration. Our approach jointly optimizes model architecture (layers, dimensions) and attention pattern. To efficiently evaluate candidates, we treat each as a pruned version of a pretrained backbone with inherited weights, thereby achieving high accuracy with minimal continued pretraining. We leverage the low cost of latency evaluation in a staged process: learning an accurate latency model first, then searching for the Pareto-frontier across latency and quality. This yields MobileLLM-Flash, a family of foundation models (350M, 650M, 1.4B) for efficient on-device use with strong capabilities, supporting up to 8k context length. MobileLLM-Flash delivers up to 1.8x and 1.6x faster prefill and decode on mobile CPUs with comparable or superior quality. Our analysis of Pareto-frontier design choices offers actionable principles for OD-LLM design.
format Preprint
id arxiv_https___arxiv_org_abs_2603_15954
institution arXiv
publishDate 2026
record_format arxiv
spellingShingle MobileLLM-Flash: Latency-Guided On-Device LLM Design for Industry Scale Deployment
Huang, Hanxian
Fedorov, Igor
Gromov, Andrey
Beckerman, Bernard
Suda, Naveen
Eriksson, David
Balandat, Maximilian
Conway, Rylan
Huber, Patrick
Sankar, Chinnadhurai
Dalmia, Ayushi
Liu, Zechun
Wu, Lemeng
Elgamal, Tarek
Sagar, Adithya
Chandra, Vikas
Krishnamoorthi, Raghuraman
Machine Learning
Artificial Intelligence
Real-time AI experiences call for on-device large language models (OD-LLMs) optimized for efficient deployment on resource-constrained hardware. The most useful OD-LLMs produce near-real-time responses and exhibit broad hardware compatibility, maximizing user reach. We present a methodology for designing such models using hardware-in-the-loop architecture search under mobile latency constraints. This system is amenable to industry-scale deployment: it generates models deployable without custom kernels and compatible with standard mobile runtimes like Executorch. Our methodology avoids specialized attention mechanisms and instead uses attention skipping for long-context acceleration. Our approach jointly optimizes model architecture (layers, dimensions) and attention pattern. To efficiently evaluate candidates, we treat each as a pruned version of a pretrained backbone with inherited weights, thereby achieving high accuracy with minimal continued pretraining. We leverage the low cost of latency evaluation in a staged process: learning an accurate latency model first, then searching for the Pareto-frontier across latency and quality. This yields MobileLLM-Flash, a family of foundation models (350M, 650M, 1.4B) for efficient on-device use with strong capabilities, supporting up to 8k context length. MobileLLM-Flash delivers up to 1.8x and 1.6x faster prefill and decode on mobile CPUs with comparable or superior quality. Our analysis of Pareto-frontier design choices offers actionable principles for OD-LLM design.
title MobileLLM-Flash: Latency-Guided On-Device LLM Design for Industry Scale Deployment
topic Machine Learning
Artificial Intelligence
url https://arxiv.org/abs/2603.15954