Start with what PrismML actually did, because the method is the story. Most small models are large models squeezed down after training, and they lose something in the squeezing. Bonsai 27B was trained in one bit from the start. Every weight is constrained to minus one or plus one, and that constraint is applied end to end, across the embeddings, the attention layers, the feed-forward blocks, and the output head, while the model learns. The result is a 27.8-billion-parameter multimodal model at a 3.9-gigabyte footprint, roughly a fourteenth of the fifty-four gigabytes a standard sixteen-bit version would demand, that still clears ninety percent of full-precision benchmark performance. A ternary variant trades a little size for more than ninety-five percent. These are not toy numbers. This is a model built to reason, write code, use tools, and read images, doing that work locally.
Now the part that matters for anyone designing a product. For the last two years, intelligence has had an address, and the address has been the cloud. Every inference is a call across the network to a rented model in someone else's data center. That arrangement is not neutral. It sets latency, because there is a round trip. It sets cost, because there is a meter running per token. It sets privacy, because the data has to leave the device to be understood. It sets availability, because no signal means no feature. Almost every real decision in an AI product traces back to that single dependency, whether the team designing it names it or not.
A capable model on the device removes the dependency. Not softens it, removes it. When the model runs locally, the work is private by default, because nothing leaves the hardware. It is instant, because there is no round trip to wait on. It is free at the margin, because there is no per-token meter. And it keeps working with the network off. Four properties that used to be in tension with capability now arrive together, and they arrive as the default rather than as a feature someone has to fight for.
When the model can run locally, the cloud stops being a requirement and becomes a choice.
Be honest about the ceiling. Bonsai 27B is not the frontier, and PrismML does not claim it is. A hosted flagship model is still smarter, still has more context to work with, still handles the hardest reasoning better. Eleven tokens a second is about reading speed, quick enough to use interactively and too slow to hand a long autonomous job. For the tasks that genuinely need the frontier, the cloud remains the right address. But a surprising share of the work real products do is not frontier work. It is classification, extraction, drafting, summarizing, routing, reading a screenshot, answering from a local corpus. That work does not need the smartest model in the world. It needs a capable one that is private, instant, and free, and that is exactly the trade Bonsai is built to make.
So on-device stops being an exotic constraint and becomes a design axis. Where the model runs is now a decision the product team gets to make, task by task, the same way they already decide which model to use. Some features route to the frontier in the cloud. Others run locally and never phone home. A product can be designed so the sensitive path stays on the device and only the hard, non-sensitive path goes out. That is a real architecture choice, and until this class of model existed it was not available. The interesting design work is drawing the line in the right place, deliberately, rather than sending everything to the cloud because there was nowhere else for it to go.
We wrote earlier in these pages that the model is rented, that the frontier is a commodity you lease by the token and that the durable value sits in everything around it. That is still true, and this is its counterweight. Renting was the only option when the cloud was the only address. Now placement joins model choice as something a team composes on purpose. The question shifts from which model do we call to which model, running where, for this task. A product that answers that question well will feel faster, cost less, and ask its users to trust it with less, not because it found a smarter model, but because it decided, deliberately, where the intelligence should live.
Bonsai 27B's architecture, footprint, and on-device benchmarks are detailed in PrismML's launch announcement.