OpenClaw and Cursor have both released native iOS apps that bring fully‑local or cloud‑based AI agents to smartphones, unlocking device capabilities, enabling remote code reviews, and signaling a shift toward ubiquitous, pocket‑sized AI assistance despite security and permission risks.
Apple’s WWDC26 introduced iOS 27 with a refreshed UI, significant performance boosts, and the rebranded Siri AI, yet regulatory constraints mean mainland Chinese users cannot access the new features, leaving them watching the showcase from afar.
At WWDC 2026, Apple introduced Siri AI powered by the new Apple Intelligence framework, showcasing context‑aware voice interactions, cross‑device integration, and a suite of iOS 27 enhancements that promise faster performance, richer UI options, and AI‑driven productivity tools across the Apple ecosystem.
The article details how the open‑source Flash‑MoE engine enables the 400B‑parameter Qwen3.5‑397B‑A17B mixture‑of‑experts model to run on an iPhone 17 Pro, achieving about 0.6 tokens per second through a custom Metal pipeline, GCD‑driven SSD streaming, and aggressive caching strategies.
OpenAI has launched a preview of Codex inside the ChatGPT mobile app for both Android and iOS, giving every user—including free‑tier members—the ability to sync code‑related tasks, approve changes, and manage remote development environments from anywhere, backed by a secure relay layer and upcoming features such as programmatic tokens and HIPAA‑compliant hooks.
Google's Gemini Intelligence upgrades Android from an operating system to an AI-driven platform, enabling cross‑app automation, Chrome‑based browsing tasks, intelligent autofill, spoken‑to‑text messaging, and natural‑language widget creation, while reshaping hardware strategy and developer interfaces.
DreamLite, a 0.39 B‑parameter diffusion model from ByteDance, unifies text‑to‑image generation and text‑guided editing in a single on‑device network, delivering 1024×1024 results in about three seconds on an iPhone 17 Pro while surpassing existing mobile and even many server‑side baselines.
Google AI Edge Gallery is an on‑device generative‑AI platform that lets users run and compare open‑source large language models on phones, offering features like chat, image Q&A, audio transcription, benchmark testing, and modular skills, while its open‑source nature and easy installation have quickly earned it over 22,000 GitHub stars.
Tencent’s Hy‑MT1.5‑1.8B‑1.25bit model, now open‑source, runs entirely offline on smartphones, supports 33 languages, and—according to internal tests—delivers translation quality that surpasses Google Translate’s online service, highlighting the impact of 1.25‑bit quantization on model size and performance.
MNN‑Sana‑Edit‑V2, a collaborative effort between Taobao’s Meta team and Hangzhou University, combines a frozen Qwen3‑0.6B LLM, Learnable Query, Connector, Linear DiT and Deep Compression Autoencoder with 4/8‑bit quantization to run fully on mobile devices, delivering 512×512 comic‑style conversions in about 15 seconds—2.5× faster than cloud alternatives—while providing open‑source code, detailed training stages, and extensive performance benchmarks.
PhoneClaw and PokeClaw are open‑source, on‑device AI agents for iOS and Android that run Gemma 4 locally, offering offline privacy, zero‑cost operation, and native tool calling through iOS APIs or Android Accessibility services.
Google’s open‑source AI Edge Gallery lets Android and iOS devices run large language models such as Gemma 4 entirely offline, eliminating network latency and privacy concerns; the app showcases six modular AI features, offers a simple install path, and signals Google’s push toward a standardized edge‑AI ecosystem.
Xiaomi’s AI team announced seven ACL 2026 papers that span low‑bit KV‑cache quantization for 8.3× faster LLM inference, OCR‑free multi‑page document VQA, a new attention‑basin analysis, non‑autoregressive spoken dialogue generation, a comprehensive mobile‑agent benchmark, a success‑rate‑aware training policy, and a progressive universal information‑extraction framework.
Google’s AI Edge Gallery lets you run cutting‑edge large language models such as Gemma 4 entirely offline on Android or iOS devices, offering absolute privacy, zero‑latency responses, and a modular platform with agent skills, thinking mode, multimodal input, and a prompt‑lab for on‑device AI experimentation.
This article presents Ada-RefSR, a novel single‑step diffusion‑based reference super‑resolution framework that introduces a "Trust but Verify" paradigm, adaptive implicit correlation gating, and lightweight architecture to robustly suppress hallucinations and achieve state‑of‑the‑art performance on multiple benchmarks, while being suitable for mobile deployment.
Liquid AI’s LFM2.5‑1.2B‑Thinking model runs entirely on a smartphone with only 900 MB of memory, scores 88 on MATH‑500, 69 on Multi‑IF, and 57 on BFCLv3 benchmarks, outperforms larger rivals, and achieves real‑time speeds on Snapdragon 8 Elite and AMD Ryzen 9 3950X, signaling a shift toward edge AI.
The article outlines the Agent‑native architecture introduced by Claude Code, detailing five core principles, file‑as‑interface design, mobile considerations, dynamic capability discovery, implementation patterns, product implications, and anti‑patterns, providing a concrete roadmap for building self‑improving AI agents.
The article explains how Tencent's open‑source HY‑MT1.5 tackles the high‑cost, large‑parameter barrier of neural machine translation by offering a 1.8 B‑parameter model that runs on roughly 1 GB of RAM, processes 50 tokens in 0.18 s, supports 33 languages, and uses on‑policy distillation to retain top‑tier accuracy, while providing a step‑by‑step online demo and free compute credits for new users.
In a live discussion hosted by Prof. Yang Jian with experts Zhang Xi and Cui Chen, the panel explores the essential abilities of mature GUI agents, the role of multimodal models in visual understanding, the transfer of code‑agent techniques to GUI tasks, edge‑device performance trade‑offs, complex planning, tool ecosystems, deployment challenges, and future breakthrough scenarios.
The article examines the underlying GUI‑Agent technologies behind the 2025 “Doubao” smartphone, comparing Microsoft’s UFO series, Alibaba’s Mobile‑Agent v2/v3, and ByteDance’s UI‑TARS, detailing their model foundations, input modalities, action spaces, planning mechanisms, learning strategies, open‑source status, and multi‑agent frameworks.
The article evaluates Doubao’s AI phone by testing everyday scenarios, highlighting its screen‑recognition‑driven automation, high latency, privacy risks, and comparing its performance and usability against Honor’s YOYO assistant and other AI‑enabled smartphones.
OPPO AI Center introduces AndesVL, an open‑source, fully‑adapted multimodal large model ranging from 0.6B to 4B parameters, designed for high‑performance, privacy‑preserving, low‑latency AI on mobile devices, with advanced architecture, training pipelines, on‑device optimizations, and state‑of‑the‑art benchmark results.
This article explains the challenges of running large language models on mobile devices, reviews recent industry efforts, and provides a step‑by‑step guide—including code snippets—for integrating a distilled GPT‑2 model with Sohu's Hybrid AI Engine using TensorFlow Lite and Keras‑NLP for on‑device inference.
TaoAvatar, Alibaba’s new 3D digital‑human platform, enables lifelike, real‑time avatars on mobile and XR devices by combining 3D Gaussian splatting, on‑device AI dialogue, and a lightweight MNN inference engine, and the full source code is now open‑sourced as MNN‑TaoAvatar.
In an interview with Vivo AI engineer Liang Tianan, the article explores the challenges of post‑Q&A recommendation, the integration of large language models into recall, ranking and evaluation pipelines, and the engineering trade‑offs required to deliver high‑quality, diverse suggestions on mobile devices.
SnapGen is a 379 M‑parameter text‑to‑image diffusion model that produces 1024 px images on mobile devices in about 1.4 seconds, using a compact U‑Net design, multi‑stage knowledge distillation, step distillation, and optimized training tricks to outperform much larger models on standard benchmarks.
SnapGen introduces a compact 379M‑parameter diffusion model that produces 1024‑pixel text‑to‑image results in about 1.4 seconds on a mobile device, achieving competitive FID scores and outperforming much larger models through a series of architecture refinements, advanced training tricks, and multi‑level knowledge distillation.
CAS‑ViT introduces a convolutional additive self‑attention mechanism that dramatically reduces the computational cost of Vision Transformers, achieving state‑of‑the‑art accuracy on image classification, object detection, and segmentation while being deployable on mobile devices.
This article explains how to integrate artificial‑intelligence capabilities such as image classification, speech‑to‑text, and facial‑expression analysis into iOS applications using Apple’s Core ML, Create ML, and Vision frameworks, providing step‑by‑step guidance, code samples, and future‑direction insights.
MNN‑Transformer provides an end‑to‑end framework that enables large language and diffusion models to run efficiently on modern smartphones by exporting, quantizing (including dynamic int4/int8 and KV cache compression) and executing via a plugin‑engine runtime, achieving up to 35 tokens/s decoding and 2‑3× faster image generation compared with existing on‑device solutions.
Google I/O highlighted the rise of on‑device AI, showing how new neural processors, Edge TPU, and tools like the Edge AI SDK and TensorFlow Lite enable developers to run large language models locally, reducing latency, cost, and privacy concerns while integrating with cloud resources.
The article outlines how Sohu’s Hybrid AI Engine enables on‑device deployment of a distilled GPT‑2 model by converting it to TensorFlow Lite, detailing the setup, customization with Keras, inference workflow, and core SDK calls, and argues that this approach offers fast, private, and cost‑effective AI for mobile devices despite typical LLM constraints.
This article explores Huolala's deployment of mobile AI image algorithms for driver document verification and vehicle sticker inspection, detailing model design, lightweighting, hybrid processing, data stream handling, and on‑device deployment that boost efficiency, privacy, and real‑time performance in logistics operations.
The talk presents an overview of the MNN inference engine, outlines the end‑to‑end workflow for deploying large language models on mobile devices, discusses technical challenges and practical solutions, and concludes with future directions for edge AI deployment.
Device intelligence brings on-device reasoning and real-time inference to smartphones and IoT gateways, delivering low-latency, privacy-preserving, personalized services such as AR/VR enhancements and recommendation re-ranking, while confronting challenges of hardware fragmentation and model size, and complementing cloud AI through architectures like Hala’s MNN-based pipeline.
The article introduces Walle, Alibaba's four‑year‑old edge‑cloud collaborative machine‑learning platform that unifies compute containers, data pipelines, and a deployment platform to enable low‑latency, privacy‑preserving, and high‑throughput AI services across billions of mobile devices, and presents its architecture, design challenges, and evaluation results.
This article explains the evolution of Ant Group's xNN edge‑AI framework, detailing its four‑layer model‑optimization space, the lightweight modeling of version 1.0, and the transition to scalable modeling in version 2.0 to better exploit fragmented device compute resources.
This article reflects on remote‑work life while introducing machine learning fundamentals and reviewing several mobile‑optimized AI frameworks—including ncnn, Caffe, TensorFlow, PyTorch, and PaddlePaddle—to help developers choose suitable tools for on‑device intelligence.
This article explains the concept of edge intelligence, its advantages over cloud‑based AI, the technical challenges of deploying AI on mobile devices, Ant Group's development timeline, core technology stack, and future directions for edge‑cloud collaboration.
The article explains how ARMv86’s new SMMLA and BFMMLA GEMM instructions are integrated into MNN to accelerate INT8 and BF16 matrix multiplication, delivering up to 90% speedup over ARMv82’s SDOT and FP16‑FMLA kernels through optimized kernels, tiling, and compatibility handling.
ShaderNN is an open‑source, sub‑2 MB GPU‑shader inference engine that runs TensorFlow, PyTorch and ONNX models directly on mobile graphics textures via OpenGL fragment and compute shaders, delivering real‑time, low‑power AI for image‑heavy tasks while eliminating third‑party dependencies and achieving up to 90 % speed gains.
Pitaya, built by ByteDance’s Client AI and MLX teams, is a comprehensive end‑side AI engineering platform that provides a full workflow from model development and data preparation to deployment, monitoring, and federated learning, supporting large‑scale commercial scenarios across multiple apps.
MNN 2.0 transforms Alibaba’s lightweight deep‑learning engine into a unified edge‑cloud framework, delivering ultra‑small binaries, broad model‑format support, and aggressive CPU/GPU/DSP/NPU optimizations—including SIMD, Winograd, quantization, and sparse computation—while providing Python‑style APIs for preprocessing, inference, and on‑device training.
To ensure consistent user experience across the wide range of Android and iOS devices, Kuaishou’s Y‑tech team designed a dynamic model‑grading framework that evaluates CPU, GPU, NPU, memory and other hardware metrics, then dispatches appropriately sized AI effect models and configurations in real time.
Kuaishou's audio‑video technology team presents a self‑developed custom video matting system that combines foreground, interactive, and video object segmentation to let creators extract arbitrary subjects without green screens, featuring adaptive cropping, multi‑stage training, and deployment across Android and iOS devices.
This article explains how Alibaba’s local‑life platforms leverage edge‑side AI to run machine‑learning inference on users’ devices, detailing the concept, advantages, technical architecture, and concrete implementations such as user feature extraction, intelligent recommendation, and smart push, while outlining future directions.
This article details the engineering optimizations applied to the image cropping stage of Kuaishou's YKit AI SDK, covering instruction-level fixes, SIMD acceleration, I/O cache improvements, algorithmic refinements, parallel processing, and device‑tier strategies to achieve up to 4.6× speedup on mobile devices.
Huawei’s 201‑million‑yuan “genius teen” Zhong Zhao leveraged AutoML to deploy high‑precision image‑pixel processing algorithms across tens of millions of Mate and P series smartphones, pioneering large‑scale commercial use of AutoML and advancing mobile visual models with dynamic convolution kernels and adversarial data augmentation.
Huawei unveiled the work of young researcher Zhong Zhao, who within a year applied AutoML to pixel‑level image processing on millions of Mate and P series phones, detailing the technical challenges, novel pipeline, performance gains, and his broader contributions to mobile AI research.
‘All‑Things AR’ (万物AR) is a Kuaishou Y‑tech solution that lets users capture any real‑world object with a phone, automatically segments it using a custom AI model, and renders an animated 3D avatar via a lightweight SLAM‑based pipeline, enabling low‑cost, high‑quality AR experiences.
The article details how to create the AI‑driven “Object Hunt Battle” game by processing data, designing and training a PP‑LCNet model with PaddleClas, converting it for Paddle.js, and integrating real‑time WebGL inference on mobile devices, achieving sub‑50 ms latency and encouraging developers to explore further.
The DeWu iOS tech salon held on September 25, 2021 brought together internal and external experts from DeWu, Alibaba, and ByteDance to share deep technical insights on video effect rendering, the MBox mobile toolchain, mobile AI with MNN, and the evolution of DeWu's iOS engineering practices, followed by interactive Q&A and community networking.
This article reviews the end‑to‑end pipeline for real‑time document corner detection on mobile devices, breaks down the keypoint detection workflow into image processing, encoding, network modeling and decoding, compares heatmap‑based and fully‑connected approaches, introduces a differentiable DSNT decoding method with unbiased coordinate transformations, and presents experimental results and conclusions on its effectiveness and limitations.
This article examines the internal structure of BERT and systematically presents various model‑compression strategies—including quantization, pruning, knowledge distillation, and structure‑preserving techniques—highlighting their impact on storage, computational cost, and inference speed for deployment on resource‑constrained mobile devices.
EdgeRec explores how moving recommendation system components to the edge—leveraging real‑time user behavior, heterogeneous action modeling, on‑device reranking, mixed‑ranking, and personalized “thousand‑person‑one‑model” training—can reduce latency, improve relevance, and boost business metrics compared to traditional cloud‑centric pipelines.
This article details the evolution of WeChat’s Identify product, explaining its end‑to‑end image recognition pipeline—including object detection, multi‑label classification, mobile‑side detection, large‑scale retrieval, unsupervised clustering, and system architecture—while showcasing various application scenarios such as product, plant, and landmark recognition.
Y‑Tech’s in‑house keypoint detection system powers Kuaishou’s beauty and effect filters across live streaming, video creation, and editing by leveraging lightweight deep‑learning models, extensive multi‑scenario data collection, and specialized handling of occlusion, enabling real‑time, robust facial and body landmark tracking on diverse mobile devices.
This week’s tech roundup highlights JD’s new federated learning platform, Facebook’s AI-driven search for renewable-energy catalysts, ARM’s high‑performance AIPU, NVIDIA’s data‑center DPU, Chrome’s rollout of HTTP/3 with QUIC, Canonical’s take on Windows‑Linux migration, plus recent advances in stereo matching and mobile sensor action recognition.
The article describes Baidu App's real-time mobile super-resolution using a VDSR-based model with pruning and depthwise separable convolutions, optimized via application-layer and inference engine techniques to halve latency and memory, enabling on-device high‑def image/video enhancement, reducing server load, and supporting iOS/Android integration.
This article reviews the motivations, challenges, and a comprehensive set of algorithmic, framework, and hardware methods—including structural optimization, quantization, pruning, and knowledge distillation—to compress and accelerate deep learning models for deployment on mobile devices, highlighting benefits such as reduced server load, lower latency, improved reliability, and enhanced privacy.
This article explains how to train and deploy convolutional neural networks directly on iOS and Android devices using Core ML, NNAPI, and TensorFlow Lite, compares performance with desktop TensorFlow, and provides practical code snippets and build‑time tips for mobile AI development.
This article explains the background of on‑device AI, introduces the Mobile Machine Learning (MML) framework and its layered architecture, details the core utilities such as model decryption and task scheduling, and provides a step‑by‑step code guide for initializing, preprocessing, inference, post‑processing, and releasing resources on mobile platforms.
The article reviews four mobile‑focused machine‑learning frameworks—NCNN, TensorFlow Lite, PyTorch Mobile (Caffe2) and FeatherKit—detailing their size, speed, and resource trade‑offs, and explains Q Music’s edge‑inference pipeline, optimization strategies, and the challenges of performance variability on heterogeneous mobile devices.
This article presents a gesture‑based human‑computer interaction system that uses Paddle Lite and MobileNet to enable real‑time control of games on Android phones, tablets, and embedded boards, detailing its architecture, data preparation, model training, and on‑device inference.
The paper presents a lightweight, anchor‑free CenterNet‑based object‑ness detector for WeChat’s Scan feature, built on a ShuffleNetV2 backbone with enlarged 5×5 depth‑wise convolutions, a streamlined detection head, and a Pyramid Interpolation Module, then quantized, ONNX‑converted and NCNN‑deployed to achieve a 436 KB model running in ~15 ms per frame on an iPhone 8 CPU.
This article details how the 58 Used‑Car mobile team built, trained, and optimized TensorFlow‑based object‑detection models for on‑device vehicle and VIN code recognition, covering data preparation, model conversion to TF‑Lite, performance improvements, engineering integration on Android/iOS, and real‑world deployment results.
Gaode’s network positioning has evolved from unsupervised clustering of massive AP fingerprints and Bayesian grid ranking to a supervised two‑level hierarchical model that scores candidate grids with a neural‑network LTR loss, while adding scenario‑specific CNN and spatio‑temporal modules for indoor, rail and subway accuracy, and it now looks toward image‑based, 5G and IoT positioning.
This article details how Alibaba leveraged AI-driven hand‑gesture detection and a lightweight SSD‑based object detection model to create an interactive rock‑paper‑scissors game for Double‑11, addressing challenges of undefined gestures, real‑time mobile performance, and data collection, and achieving over 16 million page views and high accuracy.
North Korea’s newly unveiled “Blue Sky” smartphone incorporates a homegrown AI facial‑recognition system built on CNNs, MTCNN, MobileFaceNets and TensorFlow, showcasing how the isolated nation is advancing edge AI despite operating solely on its internal CentOS‑based intranet.
This article presents Xianyu's intelligent publishing solution that leverages real‑time mobile AI to automatically associate user‑uploaded items with existing catalog entries, balancing low user cost, high accuracy, and system performance through a multi‑layer architecture and flexible pipeline design.
Intelligent publishing for Xianyu’s C2C platform uses on‑device AI to automatically match user‑posted items with the Taobao/Tmall catalog, guiding real‑time multi‑frame capture, reducing manual tagging, boosting matching accuracy by about 20%, and preparing a phased rollout for video, image, and activity posts.
Alibaba’s MNN (Mobile Neural Network) engine, now open‑sourced on GitHub, showcases how a lightweight, end‑side deep‑learning inference framework tackles fragmentation, optimizes model conversion, scheduling, and execution across diverse devices, delivering significant performance gains for mobile and IoT AI applications.
SmileAR, iQIYI’s self‑developed mobile AR platform powered by TensorFlow Lite, delivers real‑time face, body and gesture recognition across iQIYI’s apps through MobileNet‑based models, quantization‑aware training, multi‑task learning and encrypted SDKs, achieving fast, lightweight, cross‑platform AR experiences for millions of users.
This article details how Alibaba’s AliPlayStudio combines advanced computer‑vision algorithms—such as human semantic segmentation, gesture and pose detection, controllable style transfer, and face‑fusion—optimised for low‑power mobile and embedded devices, to deliver engaging real‑time video interactions across online and offline marketing scenarios.
This article details the research and engineering behind Alibaba's AliPlayStudio, a video‑interactive platform that combines computer‑vision algorithms such as human parsing, gesture and pose detection, and controllable style transfer, all optimized for real‑time deployment on low‑power mobile and embedded devices.
This article explains how Alipay’s xNN deep‑learning engine tackles the challenges of deploying AI on billions of mobile devices by using aggressive model compression, a lightweight SDK, and joint algorithm‑ and instruction‑level optimizations to achieve high accuracy, tiny package size, and real‑time performance.
The article provides an engineering‑focused guide to creating a custom TensorFlow Lite operation pipeline, covering its definition, server‑side registration and compilation, client‑side downloading, verification, decryption and dynamic loading, and discusses current limitations and possible extensions such as compression and new tensor types.
The WhereAreYou hackathon project demonstrates an iOS AR app that visualizes nearby ride‑hailing cars by converting multiple CNN models to Core ML, using ARKit to map GPS bearings, then switching to vision‑based YOLO‑style detection and tracking for real‑time vehicle identification and distance labeling.
This article explains how Alipay’s xNN engine overcomes mobile deep‑learning challenges through aggressive model compression, lightweight SDK design, algorithm‑ and instruction‑level optimizations, enabling high‑accuracy AI inference on a wide range of Android and iOS devices with minimal app‑size impact.
Mobile image search has become a critical user demand, and since its 2014 launch, Alibaba’s Paizhi Tao has evolved through multiple iterations to a robust AI-driven pipeline comprising category prediction, object detection, deep and local image feature extraction, scalable retrieval indexing, and relevance-based ranking.
Mobile visual search, a long‑standing dream, has evolved from early research to a production‑grade system at Pailitao, where a five‑module AI pipeline—category prediction, object detection, feature extraction, indexing, and ranking—enables billions of images to be searched instantly on mobile devices.
