Bonsai 27B: A 27-Billion-Parameter AI Model That Runs on a Phone

PrismML releases Bonsai 27B, the first 27B-class vision-language model compact enough to run on a modern iPhone, packing down to ~1.1 bits per weight with 256K+ token context and native tool calling.

PrismML has released Bonsai 27B, a 27-billion-parameter vision-language model that compresses down to roughly 1.1 bits per weight, making it the first model of its class able to run entirely on a modern iPhone without offloading to the cloud.

The model family includes three variants: a 1-bit Bonsai-27B at approximately 1.125 bits per weight, a Ternary-Bonsai-27B at roughly 1.7 bits per weight packed into 2-bit for accelerated kernels, and smaller 8B, 4B, and 1.7B options. All run locally via a custom llama.cpp fork, with the 27B ternary variant set as the default in the open-source demo repository.

Beyond raw compression, Bonsai 27B supports vision input including photos, screenshots, and PDFs. It offers native OpenAI-style tool calling with MCP server integration and reasoning with configurable effort budgets. Context windows stretch beyond 256,000 tokens.

The release marks a notable shift in on-device AI. Where Apple Intelligence and Google Gemini Nano operate in the sub-4-billion-parameter range, Bonsai 27B brings a model nearly seven times larger to consumer hardware through aggressive quantization. The demo repository ships with a two-command setup: clone and run, no cloud API keys required.

Sources: PrismML · GitHub Demo · Hacker News