On-Device Inference

Unified multimodal models can both understand and generate visual content within a single architecture. Existing models, however, remain data-hungry and too heavy for deployment on edge devices. We present Mobile-O, a compact vision-language-diffusion model that brings unified multimodal intelligence to a mobile device. Its core module, the Mobile Conditioning Projector (MCP), fuses vision-language features with a diffusion generator using depthwise-separable convolutions and layerwise alignment. This design enables efficient cross-modal conditioning with minimal computational cost. Trained on only a few million samples and post-trained in a novel quadruplet format (generation prompt, image, question, answer), Mobile-O jointly enhances both visual understanding and generation capabilities. Despite its efficiency, Mobile-O attains competitive or superior performance compared to other unified models, achieving 74% on GenEval and outperforming Show-O and JanusFlow by 5% and 11%, while running 6x and 11x faster, respectively. For visual understanding, Mobile-O surpasses them by 15.3% and 5.1% averaged across seven benchmarks. Running in only ~3s per 512x512 image on an iPhone, Mobile-O establishes the first practical framework for real-time unified multimodal understanding and generation on edge devices. We hope Mobile-O will ease future research in real-time unified multimodal intelligence running entirely on-device with no cloud dependency. Our code, models, datasets, and mobile application are publicly available at https://amshaker.github.io/Mobile-O/

Vision-Language-Action (VLA) models have recently demonstrated impressive capabilities across various embodied AI tasks. While deploying VLA models on real-world robots imposes strict real-time inference constraints, the inference performance landscape of VLA remains poorly understood due to the large combinatorial space of model architectures and inference systems. In this paper, we ask a fundamental research question: How should we design future VLA models and systems to support real-time inference? To address this question, we first introduce VLA-Perf, an analytical performance model that can analyze inference performance for arbitrary combinations of VLA models and inference systems. Using VLA-Perf, we conduct the first systematic study of the VLA inference performance landscape. From a model-design perspective, we examine how inference performance is affected by model scaling, model architectural choices, long-context video inputs, asynchronous inference, and dual-system model pipelines. From the deployment perspective, we analyze where VLA inference should be executed -- on-device, on edge servers, or in the cloud -- and how hardware capability and network performance jointly determine end-to-end latency. By distilling 15 key takeaways from our comprehensive evaluation, we hope this work can provide practical guidance for the design of future VLA models and inference systems.

The increasing size and complexity of modern deep neural networks (DNNs) pose significant challenges for on-device inference on mobile GPUs, with limited memory and computational resources. Existing DNN acceleration frameworks primarily deploy a weight preloading strategy, where all model parameters are loaded into memory before execution on mobile GPUs. We posit that this approach is not adequate for modern DNN workloads that comprise very large model(s) and possibly execution of several distinct models in succession. In this work, we introduce FlashMem, a memory streaming framework designed to efficiently execute large-scale modern DNNs and multi-DNN workloads while minimizing memory consumption and reducing inference latency. Instead of fully preloading weights, FlashMem statically determines model loading schedules and dynamically streams them on demand, leveraging 2.5D texture memory to minimize data transformations and improve execution efficiency. Experimental results on 11 models demonstrate that FlashMem achieves 2.0x to 8.4x memory reduction and 1.7x to 75.0x speedup compared to existing frameworks, enabling efficient execution of large-scale models and multi-DNN support on resource-constrained mobile GPUs.

Recently, Mixture-of-Experts (MoE) models have gained attention for efficiently scaling large language models. Although these models are extremely large, their sparse activation enables inference to be performed by accessing only a fraction of the model at a time. This property opens the possibility of on-device inference of MoE, which was previously considered infeasible for such large models. Consequently, various systems have been proposed to leverage this sparsity and enable efficient MoE inference for edge devices. However, previous MoE inference systems like Fiddler[8] or DAOP[13] rely on DRAM-based offloading and are not suitable for memory constrained on-device environments. As recent MoE models grow to hundreds of gigabytes, RAM-offloading solutions become impractical. To address this, we propose FlashMoE, a system that offloads inactive experts to SSD, enabling efficient MoE inference under limited RAM. FlashMoE incorporates a lightweight ML-based caching strategy that adaptively combines recency and frequency signals to maximize expert reuse, significantly reducing storage I/O. In addition, we built a user-grade desktop platform to demonstrate the practicality of FlashMoE. On this real hardware setup, FlashMoE improves cache hit rate by up to 51% over well-known offloading policies such as LRU and LFU, and achieves up to 2.6x speedup compared to existing MoE inference systems.

Recent advances in diffusion transformers (DiTs) have set new standards in image generation, yet remain impractical for on-device deployment due to their high computational and memory costs. In this work, we present an efficient DiT framework tailored for mobile and edge devices that achieves transformer-level generation quality under strict resource constraints. Our design combines three key components. First, we propose a compact DiT architecture with an adaptive global-local sparse attention mechanism that balances global context modeling and local detail preservation. Second, we propose an elastic training framework that jointly optimizes sub-DiTs of varying capacities within a unified supernetwork, allowing a single model to dynamically adjust for efficient inference across different hardware. Finally, we develop Knowledge-Guided Distribution Matching Distillation, a step-distillation pipeline that integrates the DMD objective with knowledge transfer from few-step teacher models, producing high-fidelity and low-latency generation (e.g., 4-step) suitable for real-time on-device use. Together, these contributions enable scalable, efficient, and high-quality diffusion models for deployment on diverse hardware.

Despite its efficiency, there has been little research on the practical aspects required for real-world deployment of on-device AI models, such as the device's CPU utilization and thermal conditions. In this paper, through extensive experiments, we investigate two key issues that must be addressed to deploy on-device models in real-world services: (i) the selection of on-device models and the resource consumption of each model, and (ii) the capability and potential of on-device models for domain adaptation. To this end, we focus on a task of translating live-stream chat messages and manually construct LiveChatBench, a benchmark consisting of 1,000 Korean-English parallel sentence pairs. Experiments on five mobile devices provide a systematic empirical assessment of widely adopted on-device models, highlighting the importance of model selection and deployment constraints when adapting them to specialized tasks and serving a large and heterogeneous user base. We expect that our findings will offer practical insights into both the capabilities and limitations of on-device models for real-world AI service deployment.

Camera-based animal re-identification (Animal Re-ID) can support wildlife monitoring and precision livestock management in large outdoor environments with limited wireless connectivity. In these settings, inference must run directly on collar tags or low-power edge nodes built around microcontrollers (MCUs), yet most Animal Re-ID models are designed for workstations or servers and are too large for devices with small memory and low-resolution inputs. We propose an on-device framework. First, we characterise the gap between state-of-the-art Animal Re-ID models and MCU-class hardware, showing that straightforward knowledge distillation from large teachers offers limited benefit once memory and input resolution are constrained. Second, guided by this analysis, we design a high-accuracy Animal Re-ID architecture by systematically scaling a CNN-based MobileNetV2 backbone for low-resolution inputs. Third, we evaluate the framework with a real-world dataset and introduce a data-efficient fine-tuning strategy to enable fast adaptation with just three images per animal identity at a new site. Across six public Animal Re-ID datasets, our compact model achieves competitive retrieval accuracy while reducing model size by over two orders of magnitude. On a self-collected cattle dataset, the deployed model performs fully on-device inference with only a small accuracy drop and unchanged Top-1 accuracy relative to its cluster version. We demonstrate that practical, adaptable Animal Re-ID is achievable on MCU-class devices, paving the way for scalable deployment in real field environments.

On-device machine learning (ML) has become a fundamental component of emerging mobile applications. Adaptive model deployment delivers efficient inference for heterogeneous device capabilities and performance requirements through customizing neural architectures. SuperNet-based approaches offer a promising solution by generating numerous model variants from a pre-trained ML model. However, applying SuperNet in existing frameworks requires tedious model-aware development and time-consuming hardware-aware profiling, which limits their practical adoption. We present AutoTailor, the first framework to enable automated, end-to-end SuperNet-based adaptive model deployment for edge devices. Unlike manual SuperNet construction, AutoTailor employs a computation graph-guided compilation approach to automatically transform user-provided ML models into SuperNets. To support efficient specialization, AutoTailor incorporates learning-free latency and accuracy predictors, enabling low-cost yet accurate performance prediction. Our extensive evaluations demonstrate that AutoTailor reduces the lines of code for SuperNet construction by 10.8--34.5$\times$, reduces predictor profiling time by 2.2--11.5$\times$, and achieves up to 15.60% absolute accuracy improvement and 60.03% latency reduction compared to state-of-the-art approaches across diverse models and devices. AutoTailor's source code is publicly available at https://github.com/liu-mengyang/AutoTailor.

Image- and audio-based remote IoT applications offer a substantial opportunity to address ecological changes. However, to be effective, they must operate for long periods on limited power while also providing data via LPWAN. Previous studies have investigated the energy efficiency of isolated components such as radio transmission or data processing methods within an IoT system. However, a holistic evaluation is required to determine the main drivers of energy consumption in sensor nodes. To address this, we systematically analyze energy consumption across the entire sensor-node pipeline, from data acquisition to optional on-device CNN classification and wireless network transmission, using an ESP32-S3 microcontroller. We compare on-device CNN classification followed by result transmission with transmitting raw sensor data across three environmental monitoring scenarios. Thereby, we vary developer-controllable variables, notably the model architecture and the LPWAN technologies (LTE-M, NB-IoT, LoRaWAN). Our results indicate that on-device inference can reduce sensor node energy consumption by 1.7x to 6.9x compared to raw data transmission, depending on transmission frequency, with an acceptable loss in accuracy. Further, our results narrow down the energy tradeoff to variables IoT developers can control. Therefore, our findings provide empirical guidance for designing energy-efficient, long-range sensing systems and demonstrate that TinyML enables sustainable, autonomous deployments.

Voice-based interfaces rely on a wake-up word mechanism to initiate communication with devices. However, achieving a robust, energy-efficient, and fast detection remains a challenge. This paper addresses these real production needs by enhancing data with temporal alignments and using detection based on two phases with multi-resolution. It employs two models: a lightweight on-device model for real-time processing of the audio stream and a verification model on the server-side, which is an ensemble of heterogeneous architectures that refine detection. This scheme allows the optimization of two operating points. To protect privacy, audio features are sent to the cloud instead of raw audio. The study investigated different parametric configurations for feature extraction to select one for on-device detection and another for the verification model. Furthermore, thirteen different audio classifiers were compared in terms of performance and inference time. The proposed ensemble outperforms our stronger classifier in every noise condition.