16 articles tagged
The Echo Dot went from $49.99 to $79.99 overnight. Behind it: conventional DRAM contract prices roughly 3.4x to 3.8x their 4Q25 level, and two credible forecasts pointing opposite ways.
Liquid AI released LFM2.5-DSpark speculative-decoding drafters on 20 August 2026. The headline 3.2x is a single H100 cell. On an M4 Max the 2.6B mean is 2.27x and the 8B MoE mean is 1.18x.
Apple open-sourced the Foundation Models framework utilities in 2026: Skills for just-in-time context, three history-compaction modifiers, and a chat completions client that points LanguageModelSession at any server
Google's requirement list for Gemini Intelligence names 12GB+ RAM, a qualified flagship SoC and Nano v3 or greater. Gemini Nano runs on 140 million devices. How to detect, tier and route around the gap.
How to decide where your mobile app's AI runs in 2026: Apple's on-device AFM 3 models, Google's Gemini Nano and cloud APIs compared on privacy, cost, latency and capability, with a hybrid pattern that ships.
PyTorch 2.13 landed FlexAttention on the Mac's Metal backend, with a documented ~12x speedup over SDPA on sparse masks. Here is what the numbers mean, how to write a mask in two lines, and where MLX still wins.
PyTorch 2.13 lands FlexAttention on Apple Silicon with Metal kernels for sparse prefill and decode, up to 12x faster than SDPA. How the API works, what the numbers show, and when block-sparse attention pays off on a Mac.
At WWDC 2026 Apple shipped Xcode 27 with on-device AI completion, agentic coding and a Device Hub, plus Swift 6.4 with anyAppleOS and async defer. Here is what matters for iOS developers.
Five engineering field notes from building AI features against the iOS 27 betas: the 4,096-token on-device context window, the device split, the Evaluations framework, native tools, and a moving beta target.
At WWDC 2026, Apple introduced the System Orchestrator, the iOS 27 component that decides whether a Siri request runs on-device, on Apple's Private Cloud Compute, or on Google's cloud.
Apple's Foundation Models framework grew up at WWDC 2026. The v3 AFM API adds image input, cloud routing through one call site, custom skills, and Dynamic Profiles. What changes for Swift developers.
At WWDC 2026 Apple detailed a system orchestrator that decides what runs on-device, what goes to Private Cloud Compute, and what reaches a cloud model. How the routing works, for engineers.
Apple's Foundation Models framework runs a ~3B on-device model in iOS 27. Here is the Swift protocol pattern that lets you swap to Claude or Gemini with a single line of code.
Spatial Reframing in iOS 27 rebuilds a photo's perspective using on-device spatial models and Private Cloud Compute. Four things app and product teams should know about how it works and where it fails.
Apple rebuilt its AI stack around privacy in 2026. Here are six lessons founders and engineering leaders can take from its hybrid architecture when building privacy-conscious AI systems.
At WWDC 2026 Apple introduced Core AI, the on-device successor to 2017's Core ML, and expanded the Foundation Models framework. Here are five changes iOS 27 brings for Swift developers.
