#端侧

Smart AI Keyboard: Translate

Smart AI Keyboard is an Android keyboard app offering AI writing and reply generation, grammar and spell check, tone rewriting, and real-time translation across 65+ languages, alongside swipe typing, voice input, themes, and tools such as a smart clipboard, calculator, OCR scanner, and QR scanner. It uses Android's AccessibilityService to read visible message text in supported apps like WhatsApp, Telegram, Instagram, Gmail and Messages to prepare reply suggestions, and online AI features may send submitted text to service providers. It is offered by Websoftstar Infosystems LLP, with a privacy policy at https://www.aikeyboard.in/privacy.

How to Keep an App Out of Siri AI on iPhone

iOS 27.2 added an Excluded Apps setting (Settings → Siri → Excluded Apps) that keeps chosen apps away from Siri, Search and suggestions, so Siri can't read, act in, or answer questions about content in those apps. All apps are included by default, and excluding one also removes its content from Search and stops Siri from using its data for personalization and suggestions elsewhere. The post notes Siri AI is still in beta and that an app's own access to location, photos or contacts is managed separately under Settings → Privacy & Security.

Guru: An Self Evolving AI framework for Android

GURU is a free, open-source (AGPL-3.0) Android app from Unus Lumen that acts as a permission layer across 167 Android permissions with a tool registry of around 390 tools, an encrypted Rust memory engine, a bundled Unix userland, and device control via protocols including UPnP, DIAL, Wake-on-LAN, Android TV Remote, Roku ECP, Xbox SmartGlass, SNMP, IR, SSH and ADB. It routes model calls through an embedded Tor engine that fails closed, lets users bring their own cloud, local or rented-GPU model, and can write and register its own tools, skills, automations and rooms. It is a public beta distributed as a signed ~743 MB APK (Android 8.0+) outside the Play Store, with payments and the planned data marketplace not yet live.

Show HN: EDNA – 100% on-device human sensing so AI can perceive the user

A Show HN post links to openos.space, presenting EDNA as a 100% on-device human sensing system intended to let AI perceive the user. The submission has 5 points and 1 comment on Hacker News.

Run Native DLA Inference in Python on NVIDIA Jetson AGX Orin Using cuDLA

NVIDIA describes how edge AI applications often must run multiple neural networks within limited compute and power budgets, and how the NVIDIA Deep Learning Accelerator (DLA) provides dedicated hardware for deep learning inference. On NVIDIA Jetson AGX Orin, DLA executes supported neural network operations, letting applications offload inference work from the GPU, which can improve power efficiency.

Apple HomeView available today

Apple announced that HomeView, a device that recognizes up to six people by face and shows their personal content in Messages, FaceTime, Calendar, Reminders, and Find My, is available to order today in Australia, Canada, India, New Zealand, Singapore, the UK, and the U.S. for $299 (U.S.), with a wall-mounted version also at $299. It runs Siri AI, which offers onscreen awareness and, when paired with an Apple Intelligence-enabled iPhone or iPad, personal context and cross-app actions; Siri AI is available in beta in HomeView software version 27.2. HomeView is not for sale in the European Union because Siri AI is currently unavailable in iOS in the EU, and Schneider Home mounts start at $59.95 (U.S.).

Runitlocal.io – Run AI on your own hardware

Runitlocal.io is a site about running AI on your own hardware, presented by people who run it locally. It probes your GPU in the browser without sending that data to the site, and notes that speeds are estimates unless marked as measured; users can send in their own measured results.

Edge0-145B: Frontier AI Capabilities on Your Own Computer

The company open-sourced Edge0-145B and Edge0-120B, describing Edge0-145B as the first open-source 145-billion-parameter on-device model, alongside a unified inference framework and deployment tools. It reports 29GB peak memory, 80+ tokens/s decoding, 1,100+ tokens/s prefill, roughly 120W typical power and 128K–256K context on a compatible laptop, with benchmark scores compared against named cloud models. The release also describes shared-matrix sparse experts, tiered GPU/system-memory/SSD loading and CPU-GPU co-scheduling, plus local demonstrations in video, Blender 3D scenes, games, drone control and PCB design.

What end-to-end AI systems mean for the future of automated driving

In a Q&A, Aptiv vice president of engineering Guillermo Pita explains that end-to-end AI trains a single model to learn driving behavior directly from data rather than chaining separate perception, tracking, planning and control modules, producing more human-like driving but making explainability and debugging harder. He says data diversity, not sheer volume, is the key training challenge, with simulation and synthetic scenarios used to supplement rare events, and that centralized computing architectures fit the approach while concentrating more functionality on a single platform. Success, he argues, depends on advances in AI technology and on an organization's ability to translate millions of miles of driving experience into systems that perform consistently and safely.

Local AI chat for iPhone(test build), built on a custom engine

OwnOwl is a time-limited test build that runs a large language model entirely on an iPhone, with no internet used for chatting and conversations kept on the device. It requires an iPhone with 6 GB of RAM or more and iOS 17 or newer, plus about 6 GB (GigaChat), 9 GB (Ling), or 14 GB (Gemma) of free storage for the chosen model, and is installed via AltStore or Sideloadly using your own Apple ID. Version 0.2 adds faster first replies, a Continue button for longer answers, memory of user-confirmed facts, and stability fixes; the app and its inference engine are proprietary and the build expires.

Air-LLM: Broadcasting AI Weights for Edge LLM Inference via RF Computing

The paper proposes AIR-LLM, an edge LLM inference architecture in which a central radio such as a 5G base station broadcasts LLM weights over the air and edge users perform the general matrix-vector multiplication of inference directly in the RF domain using RF mixers, avoiding weight storage and loading. It uses MIMO spatial multiplexing with an energy-efficient precoder-postcoder pair to shorten airtime, and because the central radio stays user-unaware, one broadcast can serve unlimited users in coverage. Evaluated on NVIDIA Sionna ray-traced channels of two urban scenes with a real RF mixer profile, AIR-LLM shows 4.0% WikiText-2 perplexity degradation on LLaMA-3.1-8B while saving 157.7x/40.4x energy versus FP16 and weight-only quantization baselines, and with 20 users its airtime is 104.1x/26.0x shorter.

Show HN: Senku, an AI agent that can control your phone in background

Senku is an Android AI agent, distributed as an APK rather than through Google Play, that uses Android's accessibility service to read UI elements and perform tasks in the background on a hidden screen while the user keeps using the phone. It requires permissions for display over other apps, microphone, accessibility, and wireless debugging, and it only views the screen during a task or when the user circles something; pricing is free to start, $20/month for Pro and $100/month for Max, with iOS not yet supported.

Show HN: Outlier, AI without the data center, private and offline on your Mac

Outlier is a notarized macOS app that runs local AI models entirely on Apple Silicon, keeping chats, files and code on the user's disk unless web search or a cloud model with the user's own API key is enabled. It offers six model tiers from a free 4B Nano up to a 397B Plus tier, with a one-time $49 Pro license covering all tiers and features, and requires M1 or newer on macOS 15+. The post includes benchmark tables, socket-sampling evidence of local-only inference, and an explicit list of limitations, including that it does not yet match top cloud coding agents.

Lokey – A native Android keyboard with local AI for prediction, autocorrection

Lokey is a native Kotlin Android keyboard that runs word prediction, autocorrection, dictation and translation models locally on the phone, without an account, subscription or inference service connection. It supports 39 languages and eleven keyboard families, with optional models including LFM 2.5 350M, Qwen 3.5 0.8B, Parakeet v3 and Qwen3 ASR, downloaded separately from the app. It requires Android 9 or later on an ARM64 device, is beta software, and is released under the MIT license with third-party runtimes and model weights under their own licenses.

Apple's Claris Testing 'Harbor' App for Local AI on Macs

Apple's Claris subsidiary is testing Harbor, an early-access app that turns Apple silicon Macs into a private AI inference server for open-weight models, with data staying on local hardware. Harbor can cluster multiple Macs over Thunderbolt 5 to pool unified memory for larger models and higher concurrency, using RDMA support added in macOS Tahoe 26.2, and requires no manual configuration or command line. The App Store listing has been removed, and the sign-up page asks for a company name, suggesting a business focus.

Brazil's AI Development Strategy Needs an Edge Layer – TechPolicy.Press

TechPolicy.Press argues that Brazil's AI debate is overly focused on frontier models and hyperscale data centers, while edge AI—local inference on small, specialized models—better fits the country's industrial and employment needs. The article cites Brazil's 2.3 billion reais ($444.2 million) investment in supercomputing projects with Huawei and iFlytek, and an FGV study estimating that only about 4% of jobs at a 100-megawatt data center become permanent operations and maintenance roles. It points to China's sector-specific industrial AI approach and the CESAR/UCB Power battery-inspection system in Manaus as models, and criticizes the Redata tax regime for excluding Simples Nacional companies.

BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model

Tether AI Research introduces BotPsy, a 0.8B-parameter model built on Qwen3.5-0.8B for function calling and agentic workflows, suitable for on-device deployment. It roughly doubles its base model's tool-use scores (BFCL v4 23.4 to 44.0, ACEBench 30.1 to 58.7, τ-bench 23.2 to 51.5) and outperforms every model up to 1.2B parameters tested on all six agentic benchmarks, while also beating Qwen3.5-2B-Think on five of the six. It will be released with open weights under Apache 2.0 in bf16, Q8 and Q4 as safetensors and GGUF files, and is intended for low- to medium-difficulty tasks, not high-stakes use.

ExecuTorch WebGPU: PyTorch Models on the GPU, in the Browser

ExecuTorch's new WebGPU backend runs PyTorch models exported with torch.export to a .pte file directly on the GPU in WebGPU-enabled browsers, without converting to another framework's format. The same artifact runs across Chrome and Safari on Metal, Vulkan and D3D12, and on an Apple M1 Pro in Chrome it was 4–7× faster than ExecuTorch's Wasm (XNNPACK) CPU path on prefill and 2–10× faster on decode. Validated models include Llama 3.2, Qwen3, Gemma 3, SmolVLM, Depth Anything V2, SAM 2, Whisper and on-device LoRA fine-tuning, all running locally in the tab.

Apple has new Mac app coming to make running local AI models easy

Leaker Aaron Perris found evidence of a new Apple Mac app called Harbor that turns a Mac into a private AI inference server for running open-weight models locally on Apple silicon. Harbor can connect multiple Thunderbolt 5 Macs to pool unified memory for larger models and higher concurrency, and is currently in a private preview under Apple's Claris brand.

On-Device Speech Recognition and Text-to-Speech in Google AI Edge Gallery

Google introduced on-device ASR and TTS in Google AI Edge Gallery, pairing lightweight speech models with on-device SLMs like Gemini Nano and Gemma 4 for fully offline voice interaction. ML Kit on Android offers multilingual transcription across 100+ languages at 10x real-time speed plus a few-step diffusion TTS that starts in ~100 ms, while MediaPipe Tasks and LiteRT bring cross-platform support for open-weight models including Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M, and Phonon 31M. The release powers the XR Language Teacher showcase and voice-enabled AI Chat, with open-source reference code on GitHub.

Optimizing Muse Spark Open on TPU v7 with SGLang-TorchTPU

The post describes ExecuTorch's new WebGPU backend, which lets PyTorch models exported once to a .pte file via torch.export run in any WebGPU-enabled browser through a WebAssembly runtime, covering LLMs, vision-language, segmentation, depth, detection, super-resolution, speech-to-text and on-device LoRA/DiReFT fine-tuning. On an Apple M1 Pro in Chrome, WebGPU was 4–7× faster than ExecuTorch's Wasm/XNNPACK CPU path on prefill and 2–10× faster on decode, with all data staying on the device. The backend is still in early development, and the post links an ExecuTorch WebGPU Playground and Hugging Face repos with .pte files and export recipes.

Featured coverage of this story: 《ExecuTorch WebGPU: PyTorch Models on the GPU, in the Browser》

Running AV Reasoning Models Within Onboard Compute and Latency Limits

AV developers need reasoning models that fit onboard compute and latency constraints; larger models provide a foundation, but deploying them in a vehicle requires reducing model size, preserving useful reasoning, and validating performance on target hardware. The source mentions NVIDIA Alpamayo 2 Edge in this context, but the supplied material is truncated before detailing it.

Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters

Zyphra has released ZONOS3-Mini, an Apache 2.0 text-to-speech model that is a sparse Mixture of Experts with 800M total and about 100M active parameters, supporting zero-shot voice cloning and word-level delivery control. It runs via WebGPU in the browser, MLX on Apple Silicon, ggml on CPU/CUDA/Metal/Vulkan, and SGLang-Omni on NVIDIA GPUs, with Zyphra reporting 80 ms median time to first audio on one H100 at 8 concurrent requests. Zyphra reports it ranks first among open models on its self-run ZTTS2-Eval Capability Index at 66.2 in speaker-embedding mode, while reference-clip mode scores only 44.7.

Show HN: Eigin. Local-first and no-shell personal AI for iOS

Eigin is an open-source (AGPL v3) personal AI agent for iOS whose agent runtime, memory and tools run on-device, with chats, memories and settings stored locally. It offers an on-device knowledge graph, a choice of Eigin's European cloud AI, an OpenAI-compatible provider or a local model, native calendar/reminders/health integrations plus MCP, and on-device speech recognition and text-to-speech. It is a closed-set agent with a fixed set of declared tools and no shell or virtual computer, and it asks before changing your data or acting on your behalf.

The case for and against local, personal AI

The author argues that local AI on owned hardware is now capable enough for real work and that the debate is really about who controls AI capabilities, rejecting the claim that only cloud frontier models are serious tools. He describes his own budget dual-GPU setup running models like Qwen3.8-27B and Ornith-1.5-35B-A3B, notes that RAM and VRAM prices are currently prohibitive, and predicts prices will eventually normalize. He frames a future "Windows XP era" of local AI in which families and small businesses handle private, mundane tasks without sending personal data to cloud vendors.

STMicroelectronics streamlines AI model development for STM32 developers

STMicroelectronics has redesigned the workflow in ST Edge AI Suite, adding four tools—Model Explorer, Dataset Factory, Model Factory, and Model Bench—to streamline deployment of embedded machine-learning models on STM32 microcontrollers. Model Bench replaces ST Edge AI Developer Cloud, and the company says it offers over 400 pre-trained models, with the new tools available in late November. ST says the suite supports over 260,000 projects and connects to an STM32 ecosystem containing over 4,000 commercial part numbers.

Users have started buying RTX 5060 Ti 16GB graphics cards for the sake of AI

Demand has shifted as regular users buy budget GeForce RTX 5060 Ti cards en masse for local AI workloads, driven by video memory shortages and the growing requirements of open-source generative neural networks. Some pair the 16GB card with existing GPUs such as the GeForce RTX 5070 12GB for 28GB total VRAM to run the Qwen3.8-Flash-Next IQ3_S quantized model, and enthusiasts often install two or more RTX 5060 Ti cards per system, with plans for a third via an M.2 slot using OCuLink. Using Strata version 0.2.10 with streaming KV cache data to 64GB of system RAM, a 256-token context window and roughly 246 input tokens were processed on two cards, though system RAM storage reduces pre-computation performance, so users prefer buying more accelerators.

Meta's Charm gadget carries CEO Zuckerberg's big AI ambitions

At Meta Connect on September 23, 2026, Mark Zuckerberg unveiled the Muse Charm, a pocket-watch-sized device with a roughly 2-inch screen and built-in 5G modem that lets users tap into Meta's Muse AI agent without a smartphone, part of a shift in focus toward "personal superintelligence." Zuckerberg said Muse would also soon be available across Meta's smart glasses lineup, following millions of downloads for the app. Analysts said the device could appeal to Gen Z but were cautious about mass adoption, citing the limited success of similar devices such as the Rabbit R1.

Show HN: OfflineVault – AI based offline library for when the internet is down

OfflineVault is a macOS app (Apple Silicon, macOS 14+) that bundles an offline library of emergency and reference content—a 95-article Field Guide, CDC/FEMA preparedness material, MedlinePlus, iFixit manuals, medical wikis, Wikipedia and maps—with an on-device AI assistant that answers only from the stored books and cites the passage it used. On 150 emergency questions tested through the app, the default model scored 63 fully right, 73 partly right, 4 wrong and 0 harmful, and refused all 10 questions it should refuse; every wrong answer is published. It has a free tier with no account, paid Core ($149) and Household ($249) plans, is GPL-3.0 open source, and offers ovsync for Linux and Windows via Kiwix.

Ask HN: Can anyone run frontier AI models at usable speeds?

An Ask HN post notes that eight months ago EXO labs promised fast inference on local hardware, but it is not fast enough: getting a response through hermes using Kimi K2.6 takes a minute and a half. The poster asks whether there is hardware a developer can buy that can actually offset cloud bills.

Underdog shows off a 32MB speech model in new demo

Underdog released a browser game called Woof built with its new 32MB speech generation model, which has 31 million parameters, is open for download, and the company says can run on most devices without an internet connection. The demo follows Underdog's earlier 167MB Bark speech-to-text model, and founder Sigil Wen told TechCrunch he wants to train open, American-made models that let people run capable AI locally and keep their data private. Underdog has also released a 27-billion-parameter model that can run on laptops and a waitlisted locally-hosted app.

Nanbeige has released Nanbeige4.3-33B scoring 20 on the Intelligence Index

Nanbeige, the foundation model research team at BOSS Zhipin (operated by Kanzhun Limited), has released Nanbeige4.3-33B, an Apache 2.0 open-weights dense reasoning model with text-only input and output and a 524K-token context window. It scores 20 on the Artificial Analysis Intelligence Index v5, matching Qwen3.8 27B (xhigh) at 20, with relative strengths in terminal use (30% on Terminal-Bench 4.0 vs. 6%) and physics reasoning (16% on CritPt v2's Hard Subset vs. 4%), but trails on AA-Briefcase v1.1 (1162 vs. 1391 Elo), FrontierCode 2.0 (11% vs. 15%) and HLE-Diamond (7% vs. 15%). It shows high abstention and low hallucination on Omniscience (18% vs. 29% hallucination rate, declining 71% of questions) and used 158K output tokens per Intelligence Index task vs. 110K for Qwen3.8 27B.

Aina’s smart ring has a screen, a joystick, and a mic so you can talk to AI agents

Hardware startup Aina released a $399 smart ring with a screen, joystick, mic, and speaker, letting users talk to AI agents, take notes, make calls, and control apps on Mac or phone. The company, which raised $5.5 million in July, previously made Mirage, a three-key controller for Mac apps. The ring goes on pre-order Thursday at $399 early-bird pricing, rising to $499 afterward, with shipping planned for January 2027.

A Closer Look at Phonon TTS. Launching High-Quality On-Device Speech That Tops the Benchmarks.

Gradium details Phonon, a small on-device text-to-speech model released two weeks ago that needs no server or network, with Phonon 31m (English) under Apache 2.0 and Phonon 40m covering English, French, German, Spanish and Portuguese under a non-commercial license. The models, about 65 MB and 75 MB, offer over 100 built-in voices plus voice cloning and voice design, and on Hugging Face's Open TTS Leaderboard Phonon ranks first in German and Spanish on intelligibility and tops the CPU streaming benchmark on time-to-first audio. The post also covers the ptts inference framework, Transformers integration from v5.21, launch demos, and more than 200,000 downloads in the first two weeks.

Nothing CEO says company’s mysterious AI-first device is dropping next month

Nothing CEO Carl Pei posted on X that the company's first "AI-native product" ships in December, describing it as a screenless productivity tool that "couldn't have been an app." Pei appeared to downplay links to AI wearables like pendants, pins, or glasses, saying he reads the industry's hunt for an "iPhone moment" differently. No firm release date has been announced.

Advancing the Quality-Efficiency Frontier for Coding AI on Windows PCs

Microsoft describes MAI-Code-1.1-Flash, a 137.9-billion-parameter mixture-of-experts coding model quantized from roughly 276 GB at 16-bit precision to a 53.0 GiB mixed-precision checkpoint (with a 46.1 GiB variant), running locally on NVIDIA RTX Spark PCs via GitHub Copilot. Using a DFlash2 speculative-decoding draft with layer-wise sliding-window attention, the measured RTX Spark configuration reached a median 51.12 tokens per second at 120K-token input context, with median end-to-end latency falling 27% from 527.25 to 382.41 seconds.

Kenyan startup builds AI around mobile money and low-cost phones

Fastagger, a Kenyan startup founded by Mutembei Kariuki, Jude Mwenda and Stephanie Njerenga, builds AI models that run offline and on low-power cellphones; its first major product, Auni, a business-intelligence assistant launched in 2024 and integrated into M-Pesa's business app, has signed up about 6,000 businesses. Auni helps small businesses such as barbershops and restaurants analyze transaction data and get marketing recommendations, with some functions like reading sales statements and visualizing results running on-device without cloud queries. Fastagger has support from Google, Microsoft and Nvidia accelerator programs but still has just its three founders, and is working on adding agents that could automatically implement marketing actions.

My therapist thinks I should chuck the Amazon Bee Frontier out a window

In this Optimizer newsletter, Verge senior reviewer Victoria Song recounts three weeks of wearing the always-recording Amazon Bee AI wearable, including therapy and medical appointments, and describes its transcripts and summaries misremembering conversations, inventing six extra cats, and generating blunt summaries of a childhood trauma. Her therapist argued such a device supersedes a person's ability to trust their own instincts and suggested she throw it out a window. Song also criticizes Bee's Grow feature for judging her hobbies as lacking "shareable output."

Nothing teases a screen-free AI device ahead of launch next month

Nothing CEO Carl Pei said in a post on X that the company will launch its first AI-native product in December, describing it as a tool with no new screen that "couldn't have been an app." He shared an image reading "NOTHING(R)" above an ambiguous black rounded corner, and said that "looking for the iPhone of AI today is a dead end." The Verge notes earlier rumors that Nothing plans two pairs of AI smart glasses, including one without cameras, and is repositioning as an AI-first company.

I tried Apple’s Siri Recaps. They’re surprisingly useful — and a little unsettling

Apple released Audio Intelligence features in the first developer betas of iOS 27.3 and watchOS 27.3, including sound detection, song identification, Live Rewind, and Siri Recaps, with audio processed on-device where possible and without storing raw audio or identifying speakers. In a day of testing, the author found Siri Recaps captured useful details like a repair deductible, time frame, and rental gas level, but also saved extraneous items such as TV plans and mixed unrelated topics into one note. The author suggests separate notes or sections for extraneous details, swipe-to-delete on individual bullet points, and an option to auto-export saved Recaps to Apple Notes.