PrismML, a startup focused on model compression, recently unveiled an impressive achievement: Bonsai 27B, the first 27-billion parameter large model that can run on a smartphone.

This model is the new multimodal flagship of the Bonsai family, trained based on Qwen 3.6 27B. According to official statements, it can achieve multi-step reasoning, structured tool calling, long-context workflows, and coherent agent loops entirely on local edge devices—capabilities that were completely out of reach for previous small edge models.

Before this breakthrough, 27B-scale models could never be deployed locally. At standard 16-bit precision, a 27B model requires roughly 54GB of memory, and even when compressed to 4-bit, it still needs around 18GB. That is too large for most consumer laptops, let alone smartphones.

Bonsai 27B changes this situation through extreme quantization.

It comes in two versions:

1. Ternary quantized version: total size of 5.9GB, with an average of 1.71 effective bits per weight, optimized to deliver usable output quality on laptops

2. 1-bit quantized version: compressed to 3.9GB, with an average of 1.125 effective bits per weight, specifically tailored to fit smartphone storage capacities

Independent testing by netizens shows the 1-bit version requires less than 4GB of running memory, and it can run smoothly on iPhone 15 Pro and the latest flagship Android devices. Its file size is even smaller than the installation package of many popular AAA mobile games.

This project is backed by Khosla Ventures, a high-profile edge AI startup in Silicon Valley. The entire project, including all model weights, is now open-source on GitHub under the Apache 2.0 license, which imposes almost no restrictions for both personal and commercial use.

A year or two ago, few people could imagine that a 10-billion-plus parameter large model could fit right in your pocket. Edge AI is advancing much faster than most industry insiders predicted. Now we are just waiting for real-world test results to see how well the model actually performs after trading off precision for smaller size.