PrismML Releases Bonsai 27B, a 27B-Parameter AI Model That Runs on iPhone 17 Pro
The release demonstrates that a 27-billion-parameter model can be compressed to fit within a phone's memory while preserving most of its original capability, making on-device inference practical for offline, private use.
Reporting from 1 source: GIGAZINE.
PrismML released Bonsai 27B, a compressed version of Qwen3.6-27B that runs on iPhone 17 Pro. The model comes in two variants: a 1-bit version at 3.9GB that retains about 90% of the original's performance, and a ternary version at 5.9GB that retains about 95%. It handles up to 262,000 tokens of context, supports multimodal inputs and tool calling, and is published under the Apache 2.0 license.
The 1-bit variant uses only 3.9GB of memory, placing it within the app-accessible memory limit on iPhone 17 Pro. The ternary variant uses 5.9GB. On a 15-benchmark aggregate, the original Qwen3.6-27B scored 85.0; the ternary version scored 80.5, and the 1-bit version scored 76.1. PrismML reports the ternary variant maintains about 95% of the original model's performance and the 1-bit variant about 90%.
Bonsai 27B supports a context window of roughly 262,000 tokens and accepts text, screenshots, documents, and camera images as input. It can call external tools and handle agent-style workflows. Under Apache 2.0, the model weights are available on Hugging Face in MLX (Apple) and GGUF (llama.cpp) formats.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.