CourionAI Newsletter
← All news
open-source 2 min read

A Capable AI Model That Fits on Your Phone, and Runs Entirely Offline

Startup PrismML shrank a 27-billion-parameter open model to under 4 GB so it runs on an iPhone, keeping about 90 percent of its ability. Apple is reportedly testing the compression tech. Why on-device AI is quietly a big deal.

Risograph illustration of a bonsai tree growing out of a smartphone with roots spreading into circuit-board traces, evoking a powerful AI shrunk to fit in your pocket

Most AI you use today lives in the cloud: you type, your words travel to a data center, and the answer travels back. A startup called PrismML, founded by a group of Caltech researchers, just showed a different path. Its new model, Bonsai 27B, is a full-featured AI (it can reason through multi-step problems, use tools, and read images) that runs directly on an iPhone, with no internet connection needed. The model weights are open under the Apache 2.0 license, so anyone can download and run them.

The clever part is the shrinking. A model this size would normally need around 54 GB of storage, or about 18 GB even after the usual squeezing. PrismML got it down to roughly 3.9 GB, small enough to fit on an iPhone 17 Pro Max. They do it with aggressive “quantization,” which just means storing each of the model’s internal numbers with far less detail (in the most extreme version, each value is boiled down to one of only two states). In the company’s own tests across 15 benchmarks, the phone-sized version kept about 90 percent of the full model’s ability, and its math and coding skills were “virtually unaffected.” On an iPhone it produces about 11 words per second, and a full battery is good for roughly 67,000 words before it drains.

What’s behind it: As one caveat, these are PrismML’s own numbers, so independent testing will tell the real story, and the heaviest compression did visibly hurt some skills like reading images and following fiddly instructions. But the direction matters. Modern AI “agents” (tools that work through a task step by step on their own) can make hundreds of calls to a model for a single job. In the cloud, each call costs money, adds a delay, and sends your data (screenshots, documents) off your device. Running on your phone makes those loops essentially free and keeps your data with you. That is why this is interesting for Apple in particular: its own on-device AI has trailed rivals, and CNBC reports PrismML is already in early talks with Apple about licensing the compression, though the company calls them “very early.”

What this means for you: Nothing to install today unless you like tinkering (there is a free demo and a developer preview). But this is the clearest sign yet that “private, offline AI on your own phone” is becoming real rather than a niche hobby. For anyone who worries about where their prompts and documents end up, that is genuinely good news. Expect the phones and laptops you buy in the next couple of years to lean on this kind of on-device AI for the everyday stuff, and save the cloud for the hardest tasks.

Sources

Source: https://prismml.com/news/bonsai-27b

Next story

Google Search Will Now Make Up an Image When It Can't Find One

Google is adding AI image generation to Search: when no matching picture exists on the web, you can type a prompt and it creates one. Powered by a fast, cheaper 'Nano Banana 2 Lite' model. What it means for you, and for the web.

Risograph illustration of a search bar with a magnifying glass and a paintbrush conjuring a picture into an empty frame, evoking a search engine inventing an image it cannot find