---
title: "We are partnering with PrismML"
description: "PrismML compresses open models down to a fraction of their normal size, which means bigger models now fit on a phone. Here is what that means for Offeline."
pubDate: 2026-08-20
author: "Offeline"
---


We are partnering with [PrismML](https://prismml.com), and the short version is that bigger models are about to fit on your phone.

Here is the problem we keep running into: phones are not short on ambition, they are short on memory. A normal 8B model at the usual 4-bit quantization lands somewhere around 4.5GB, and iOS only hands an app roughly half the phone's RAM. On an 8GB phone that means about 4GB to work with, and the model has to share that space with the app and with the conversation as it grows. So the good models sit just outside the door on most phones.

PrismML goes at this from the compression side. Their Bonsai family ships in 1-bit and ternary builds instead of the usual 4-bit, and the difference is not small. Bonsai 8B is built on Qwen3-8B and carries 8.19B parameters, but the 1-bit GGUF build is about 1.15GB against 16.38GB at full precision. That is roughly a fourteenfold reduction. Bonsai 4B goes smaller again at about 0.57GB for the 1-bit build and 1.07GB for the ternary one.

Read those numbers against the memory math above and you can see why we got interested: an 8B class model at 1.15GB fits on a 6GB phone with room left over, where the 4-bit build of the same model never had a chance. It moves a model between hardware tiers, which is a nicer way of saying it moves a good model onto a phone somebody already owns.

Both sizes are open weights under Apache 2.0 and both run on llama.cpp and MLX, so there is no licensing catch and nothing proprietary you have to opt into. That mattered to us, because Offeline runs open models by choice and we did not want a partnership that quietly changed that.

If you want the wider picture on quantization and what fits where, our [guide to local AI models](/blog/complete-guide-local-ai-models-2026/) covers it, and [which model your iPhone can run](/blog/which-local-ai-model-can-your-iphone-run/) has the memory tiers laid out. For the compression work itself, PrismML have their own writeups over at [prismml.com](https://prismml.com).