PrismML Bonsai 27B: AI Model Runs on Your Phone

PrismML has launched Bonsai 27B, a new AI model designed to run locally on smartphones and laptops. This compressed model, based on Qwen3.6 27B, uses about 3.9 gigabytes of memory for its 1-bit version. It supports coding, reasoning, image understanding, and tool use. What's interesting is that it can process both text and images without needing to send every request to a cloud server. The 1-bit version reportedly generates around 11 tokens per second on an iPhone 17 Pro. The model achieves its small size by storing model weights using only one bit. This is a significant reduction compared to most AI models. PrismML also offers a ternary version, which uses three possible values for each weight, aiming to preserve more performance. Benchmark scores show the full-precision Qwen model scored 85 overall. The ternary Bonsai 27B scored 80.5, while the 1-bit version scored 76.1. Math performance remained relatively close, but there was a bigger drop in tool use and vision scores for the compressed versions. The bottom line is that running AI locally on a phone can reduce reliance on an internet connection and keep personal data on the device.