The most powerful AI models have always had one thing in common: you needed serious hardware — and serious money — to run them. PrismML just binned that assumption entirely.
Their new model, Bonsai 27B, is a full 27-billion-parameter reasoning AI that runs locally on an iPhone. No cloud subscription. No monthly fee. No sending your data off to some server farm in Virginia. Just your phone, doing the heavy lifting, for free.
We've seen plenty of "on-device AI" claims before, but they've almost always meant something stripped-down — a lightweight model that can autocomplete a sentence or summarise a paragraph. Bonsai 27B is not that. It's a reasoning model. The kind of AI that thinks through problems step by step rather than just pattern-matching its way to a plausible-sounding answer. Running that class of model on a smartphone is a genuinely different thing.
What Makes This Actually Significant
The size matters here. Twenty-seven billion parameters is not a small model. For context, the race to shrink capable AI down to consumer hardware has been grinding on for years, and the general assumption has been that proper reasoning capability required you to be online and connected to something much bigger. Bonsai 27B breaks that ceiling in a way that hasn't been done before at this scale.
PrismML have made it free to use, which is the other part of this worth paying attention to. The AI industry has spent the last couple of years locked in a subscription war — everyone from OpenAI to Anthropic is charging monthly for access to their better models. The question of who controls the infrastructure, and who pays for it, shapes everything about how these tools get used. An on-device model that costs nothing changes that dynamic. Your data stays on your device. The compute is yours. Nobody is metering your usage.
This sits in an interesting moment for the broader AI landscape. [Claude's personality reportedly shifts depending on which model version you're using and even the language you speak](/getohedz/crypto/claude39s-personality-changes-depending-on-the-modeland-the-language-you-speak) — which is a reminder of how much the centralised model approach puts users at the mercy of decisions they can't see or control. Local AI sidesteps that entirely.
The Bigger Picture
We're watching a genuine split develop in how AI gets deployed. On one side, the big labs are building ever-larger cloud models and charging for access. On the other, a quieter movement is pushing capable models down to the hardware people already own. Bonsai 27B is the clearest statement yet that the second approach is further along than most people realised.
The AI girlfriend apps raking in [hundreds of millions in subscription revenue](/getohedz/crypto/this-is-how-much-your-boyfriend-is-spending-on-ai) are a decent illustration of where the subscription model ends up — monetising engagement and dependency. A free, local reasoning model points somewhere different.
Our take: Bonsai 27B is not a prototype or a proof of concept. It's a working 27B reasoning model running free on consumer hardware, right now. PrismML have quietly done something the bigger players said wasn't ready yet. Whether the industry takes that seriously is another question — but we'd suggest they probably should.