Apple AFM Plus 150B spotted
- Leaked screenshots of Apple’s internal AFM Playground app showed unreleased “AFM Plus 150B Instruct” and “AFM Plus 30B Instruct” model options on May 1. (testingcatalog.com) - The standout detail is the 150B label — far above Apple’s public ∼3B on-device model and consistent with a much larger cloud-side system. (machinelearning.apple.com) - That matters because Apple has publicly exposed only its on-device model and a server model — not a named 150B “AFM Plus” tier. (machinelearning.apple.com)
Apple’s AI story just got a little more interesting. Screenshots shared this week appear to show an internal Apple app called AFM Playground, and inside it are mode(testingcatalog.com)hots are real, the takeaway is simple: Apple is testing a much bigger in-house model stack than the one it talks about in public. (testingca([machinelearning.apple.com)at is AFM Playground? It looks like an internal chat-style testing app for Apple Foundation Models — basically a sandbox where empl(machinelearning.apple.com)ough and utilitarian, which is exactly what you’d expect from a benchmarking tool rather than a consumer app. (testingcatalog.com) ### What exactly leaked? The important part is the model picker. That’s where the screenshots reportedly show two unreleased entries: “AFM Plus 150B Instruct” and “AFM Plus 30B Instruct.” “Instruct” usually means a model tuned to follow user requests cleanly, more like ChatGPT than a raw base model. (testingcatalog.com) ### Why is 150B a big deal? Because Apple’s public messaging has centered on two things: a roughly 3-billion-parameter on-device model and a larger server model running through Private Cloud Compute. Apple has described the newer server model as a m(testingcatalog.com)r a new internal naming layer, a new generation, or both. (machinelearning.apple.com) ### Does this mean Apple built a frontier model? Maybe — but don’t overread it. A 150B label tells you the system is large, but it do(testingcatalog.com)rained versus adapted from another training pipeline. The leak is evidence of internal testing, not proof of public deployment or product readiness. (testingcatalog.com) ### Why have a 30B model too? That part may be just as revealing. A 30B tier sits in an awkward middle ground — too big for phones, but potentially useful for(machinelearning.apple.com)adder of models instead of one giant cloud brain and one tiny phone model. (testingcatalog.com) ### How does this fit Apple’s public AI strategy? Pretty neatly, actually. Apple already says its foundation models come in on-device and server forms, and that the latest versions improve tool use, reasonin(testingcatalog.com)ls Apple uses to push Siri and Apple Intelligence beyond the lightweight version developers can access. That last part is inference, but it lines up with the architecture Apple has already described. (machinelearning.apple.com) ### Is this the rumored future Si(testingcatalog.com)ence, but that’s still speculation. A testing harness can resemble a product without being a product — like seeing the backstage control panel and assuming it’s the show. (macmagazine.com.br) ### So what should you actually take from this? The leak does not prove Apple is about to ship a 150B assistant. But it does strongly suggest Apple is experimenting with larger proprietary models(machinelearning.apple.com)ion-tuned variants. That’s the real signal. Apple’s public AI stack may still look cautious — but internally, it seems much less small-ball. (testingcatalog.com)
