For three years the AI industry has run on a simple assumption: the smarter you want a model to be, the bigger it has to get — and the bigger it gets, the further away it has to live, humming inside someone else's data centre. A young lab called PrismML is quietly arguing that assumption is wrong, and it now has a model small enough to make the argument stick.

PrismML has not raised the kind of money that buys headlines. Its seed round sits at $22.25 million — a rounding error next to the war chests of the frontier labs. What it has instead is a specific technical thesis and the credentials to pursue it. The company was started by a group of Caltech researchers and is run by Babak Hassibi, a Caltech professor whose academic career has centred on compression. Ion Stoica, a co-founder of Databricks and director of Berkeley's Sky Computing Lab, advises the company. Khosla Ventures, Cerberus Capital and Caltech are on the cap table.

Its latest release, Bonsai 2 27B, takes Qwen3.8 27B — a widely used open-source model from Alibaba — and compresses it down to 5.9 GB. That is a nine- to ten-fold cut in memory footprint, and small enough to sit comfortably on a laptop and, plausibly, on a high-end phone. Speculation that the lab is in conversation with Apple has been circulating; the CEO has not commented on it.

Shrinking a model is not hard. Shrinking one without lobotomising it is. Plenty of compression techniques trade away accuracy until the resulting model is fast, tiny and not worth using. PrismML's pitch rests on how little it appears to give up: Bonsai 2 lands at 98% of the original model's aggregate benchmark scores, up from 95% for the first Bonsai release. Two releases, three points of recovered performance — a trajectory that matters more than either number on its own.

Hassibi is candid that some loss is probably permanent — compression will always cost something. But the remaining gap is arguably academic. Uncompressed models are themselves imperfect, and benchmark scores are a rough proxy for real-world usefulness at best. A two-point difference on a leaderboard rarely translates into a difference a user would notice. The software harness a model runs inside has repeatedly been shown to influence accuracy at least as much as the weights themselves.

The technique comes down to how the model stores what it knows. A model's weights are the values it absorbs during training, and conventionally each one is stored using 16 bits of precision. PrismML's approach, which it calls ternary weights, collapses that range to just three possible states: +1, −1 or 0. Storing three options instead of thousands of gradations means each weight consumes a fraction of the space, and the total footprint falls off a cliff.

The obvious objection is that throwing away that much precision should wreck the model. The lab's results suggest it does not — at least not by much — and that is the whole commercial proposition.

The lab's next target is far larger: models in the several-hundred-billion-parameter range. Counterintuitively, Hassibi expects that to be the easier problem. Bigger models carry more redundancy, which leaves more slack to squeeze out before intelligence starts to degrade. If that holds, full performance parity becomes more achievable as models grow, not less — the opposite of what the scaling-is-everything narrative would predict.

Adoption suggests developers are already convinced enough to try it. The first Bonsai model has been downloaded more than 11 million times, with a further 2.6 million downloads across the lab's smaller variants. For a company most people have never heard of, that is meaningful distribution.

The strategic question for founders is what a world of capable local models does to the assumptions baked into today's AI products. A great deal of current infrastructure — API billing, usage tiers, inference margin, cloud lock-in — exists because intelligence has to be rented. If a sufficiently good model fits on the device in a buyer's pocket, that whole layer becomes optional for a large class of applications.