Apple spotted AFM Plus 150B

Apple’s AI story just got a little more concrete. An internal Apple app called AFM Playground has surfaced with model options labeled “AFM Plus 150B Instruct” and “AFM Plus 30B Instruct.” That matters because Apple has already told the world it runs a compact roughly 3B-parameter model on device, plus a larger server model through Private Cloud Compute. What changed now is the naming — and the scale. It looks like Apple is testing a fresh “AFM Plus” family right before WWDC. (testingcatalog.com) ### What is AFM Playground? It looks like an internal chat-style testing app for Apple Foundation Models — basically a harness where Apple employees can swap models, try prompts, and compare behavior. The interface shown so far looks rough, more like a lab tool than a product. That’s important because it suggests evaluation and tuning, not a consumer app leak. (testingcatalog.com) ### Why is the 150B label such a big deal? Because 150B is a completely different class from the model Apple has publicly emphasized for on-device use. Apple’s own research pages describe an approximately 3B-parameter on-device model as the core local model behind Apple Intelligence. A 150B model is far too large for (testingcatalog.com)ack. (machinelearning.apple.com) ### Didn’t Apple already have bigger server models? Yes — but Apple has mostly talked about them in broad terms. In 2024 it described a larger server-based model available through Private Cloud Compute, and in 2025 it said its latest lineup still paired the on-device model with a server model optimized for Apple silicon (machinelearning.apple.com)second 30B tier that sits in the middle. (machinelearning.apple.com) ### So what does “AFM Plus” imply? The simplest read is “new generation.” TestingCatalog notes that the “Plus” label is new, and that’s the interesting part. Apple already uses AFM shorthand for Apple Foundation Models in the broader ecosystem. Adding “Plus” to 30B and 150B variants suggests Apple may be separating its next wave of models from the first Apple Intelligence generation instead of quietly revising the old stack. (testingcatalog.com) ### Why include a 30B model too? Because a middle tier solves a real deployment problem. A 30B model is still too heavy for phones, but it’s much easier to run than 150B. That makes it a plausible fit for Macs with stronger Apple silicon, faster cloud responses, or narrower tasks where Apple wants better quality than (testingcatalog.com)t fits the architecture Apple has already described. (testingcatalog.com) ### How does this connect to WWDC? WWDC is where Apple has been laying out its AI stack in public. In 2024 it introduced Apple Intelligence and the split between local and server models. In 2025 it expanded that story with updated foundation models and a developer-facing Foundation Models framework. So a leak showing (testingcatalog.com)what comes next. (machinelearning.apple.com) ### Is this proof Apple will ship a 150B model? No — and that’s the catch. Internal selectors are evidence of testing, not of product plans. Companies evaluate lots of models that never become public features. But this is still meaningful because the names line up with Apple’s known architecture, and they show Apple experimenting with scale beyond the small on-device model it has centered in marketing. (testingcatalog.com) ### What’s the bottom line? The leak doesn’t prove a launch. But it does make Apple’s AI roadmap look less like one tiny local model plus vague cloud magic, and more like a real model ladder — 3B on device, something mid-tier around 30B, and a flagship-class 150B system behind the scenes. If Apple wanted WWDC to show (testingcatalog.com)ect to surface first. (testingcatalog.com)