Tag: #on-device36 results
PocketBot is a fully local on-device AI agent for iPhone that converts English to automations using a quantized 3B model via llama.cpp on Metal. The team seeks advice on better sub-4B models for tool calling, optimal quantization like Q4_K_M vs Q5_K_S, dynamic sampling, and context management. Beta is available on TestFlight.
Qualcomm launched the Snapdragon Wearable Platform Supreme Edition, fitting 2 billion AI parameters into a brooch-sized device. This release completes the final puzzle piece in Qualcomm's personal AI ecosystem for wearables. It enables powerful on-device AI processing in compact form factors.
Solo developer launches Speaklone app enabling on-device Qwen3-TTS inference (1.7B on macOS, 0.6B on iOS) via MLX-Swift, with voice cloning from 5-30s samples and natural language voice design. It supports streaming TTS without cloud dependency, overcoming memory constraints through quantization, cache limits, and chunked decoding. Now available on the App Store.
Co-founders of Reface and Prisma launched Mirai to enhance on-device AI model inference. The startup secured $10M seed funding. Targets optimization for smartphones and laptops.
Ruva introduces a 'Glass Box' architecture for Personal AI using Personal Knowledge Graphs, allowing users to inspect and precisely redact AI knowledge. It replaces vector databases with graph reasoning to eliminate hallucinations, ensure accountability, and enable the 'Right to be Forgotten.' A demo and project are available online.
Same INT8 ONNX model tested on 5 Snapdragon chipsets yields accuracy from 93% (8 Gen 3) to 71% (4 Gen 2), vs 94% cloud. Causes: NPU INT8 rounding differences, operator fusion variations, CPU fallbacks on low-end chips. Highlights need for hardware-specific on-device testing.
Apple's iOS 27, iPadOS 27, and macOS 27 will expand Apple Intelligence with new AI-powered photo editing tools like Extend, Enhance, and Reframe. These features allow background expansion, automatic quality improvements, and perspective shifts for spatial photos. The updates aim to match competitors like Google and Samsung, with WWDC reveal expected June 8.
Nothing introduces a new on-device AI-powered dictation tool. The tool supports over 100 languages for seamless input.
Voice AI is expanding beyond call centers to sites like vending machines and funeral homes. Lightweight small language models (SLMs) enable net-free on-device AI. Japan can win globally by combining hardware excellence, anime IP, and omotenashi spirit.
User praises Gemma 4 E2B as daily driver on Pixel 10 Pro via AI Edge Gallery app, using CPU acceleration at 32K context. Feels 'ridiculously intelligent' like a 7B model, faster than reading speed with function calling. E4B too slow on phone.