Bonsai 1.7B is an open model by Prism ML: built from Qwen3-1.7B and trained end-to-end for 1-bit weights, it shrinks to 0.24 GB — small enough to run entirely in your browser on WebGPU. This demo turns it into a support-message triage assistant: paste a message and it returns category, urgency and a one-line summary. Nothing you type leaves your device. A sysf.io demonstration of sovereign, on-device inference.
1.7B PARAMS · 1-BIT Q1_0 (≈1.125 BPW) · 0.24 GB · 32K CONTEXT · APACHE-2.0
MODEL © PRISM ML (APACHE-2.0) · ENGINE: WEBML-COMMUNITY (MIT), PORTED TO QWEN3 BY SYSF.IO · FREE-CHAT VARIANT
Enter a Hugging Face access token with read access to prism-ml/Bonsai-1.7B-gguf. It stays in this browser — requests go only to huggingface.co.
Bonsai-1.7B reads it on your GPU and returns category, urgency and a one-line summary. Each message is triaged independently. Nothing leaves your device.
Kernels are the low-level GPU programs that do the model's actual math — the 1-bit matrix multiplies, attention and normalizations behind every token. Each one is a WGSL compute shader, compiled by your browser for this GPU.