Your AI vendor is hoping you don't read this Last week a 27-billion-paramet...

Your AI vendor is hoping you don't read this Last week a 27-billion-parameter model got small enough to run on a phone. For free.

PrismML, a Caltech spinout, took Qwen3.6 and crushed it down to 3.9 GB (yes, small enough for an iPhone). Apache 2.0, so you can actually run it. That's a workstation-class model living on the laptop already on your desk.

Here's why a business owner should care (and where I'd pump the brakes).

No API meter running. Your data never leaves the building. No vendor holding the off switch. This is what "own your AI" looks like on a small-business budget, and last year the same capability needed a server rack.

Now the part the demos skip. Compression isn't free. Tool-calling drops hard (one independent test fell from 80 to 66), and every compressed build went 0-for-9 on long, multi-step workflows. So this is a single-shot tool (summarize this, sort that, pull the fields out of a document), not an agent you turn loose.

The 5.9 GB version is the one to take seriously. The 3.9 GB one is a great party trick.

In plain English: great for the boring, high-volume text jobs you'd never want to pay per-token for. Not ready to run the business on its own.

The signal isn't this one model. It's how fast the gap between "frontier model" and "runs on your phone" is closing, a lot faster than the pricing pages suggest. So next time you're about to sign a per-seat AI contract for something as simple as summarizing documents, ask whether you could just own the thing instead. Increasingly, you can.