Apple teases on-device AI health coach

Apple’s health story is getting more interesting — and more constrained. The original idea was a big AI health coach inside a revamped Health app, something that could look at your Apple Watch and iPhone data and give personalized guidance. But by February 2026, that plan had been scaled back, with Apple reportedly winding down Project Mulberry as a standalone push and deciding to ship pieces of it more gradually inside the Health app. That matters because Apple has spent years building the raw ingredients for exactly this kind of product. The company already frames machine learning around on-device use across iPhone, Apple Watch, and Mac, and its developer materials keep stressing local inference rather than cloud-first AI. On the health side, earlier reporting said the coach was meant to use data Apple already collects — things like activity, sleep, and other health signals — and mix that with guidance shaped by physicians and health content. So what changed? Basically, the ambition ran into reality. A general-purpose “AI doctor” is a much harder product than a bundle of narrower coaching features. Health advice is high stakes. It has to be accurate, careful, and easy to explain. If Apple keeps more of that work on-device, the constraints get even tighter — less memory, less room for giant models, and less tolerance for latency or battery drain. That makes a piecemeal rollout make sense. The second half of the story is the hardware trend underneath it. Apple’s MLX framework has become a real center of gravity for local AI on Apple silicon. MLX is built around Apple’s unified memory model, which lets models and arrays move more naturally across CPU and GPU without the same kind of copying overhead you hit on more fragmented systems. That does not remove the memory problem, but it changes the tradeoffs enough that developers keep trying bigger local models on Macs. That is where Gemma 4 enters. Google released Gemma 4 on March 31, 2026 in E2B, E4B, 26B A4B, and 31B sizes, pitched explicitly as a family meant to run on your own hardware. Google also says the models support up to 256K tokens and more than 140 languages, which is impressive but comes with a catch — large context windows and larger variants demand a lot more memory. In plain English, yes, you can run serious models locally now, but the machine still decides what “serious” means. You can already see developers pushing that edge on Macs. There are MLX-based and Apple-silicon-focused projects packaging Gemma 4 for offline use, including builds aimed at local servers and OpenAI-compatible APIs. Some of them promise fully offline operation, private data handling, and decent performance on M-series Macs — but they also spell out the hardware floor. A lightweight Gemma 4 setup can fit on a 16 GB machine, while larger variants start pushing into 32 GB territory and beyond. That is the real connection between Apple’s health coach tease and the Gemma-on-Mac demos. The story is not just “AI goes local.” The story is that companies now want AI features that stay private, stay responsive, and keep working when the network is bad — but they have to design around memory ceilings, thermal limits, and model size. Health is the clearest example because the privacy upside is obvious and the risk of bad advice is obvious too. The bottom line is simple. Apple still looks committed to AI in health, but not as one giant magic coach dropped all at once. The likely path is smaller, narrower, hardware-aware features first — the kind of AI that can run locally, fail gracefully, and earn trust before it tries to sound like your doctor.