Bonsai-1.7B-unpacked

Bonsai-1.7B-unpacked

FP16 safetensors copy of the 1-bit Bonsai-1.7B model, for running it in standard Hugging Face tools. The card advises the native 1-bit releases instead, which are smaller and faster.

Made by prism-ml on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This repository holds Bonsai-1.7B, a model of about 1.7B parameters, unpacked from its 1-bit form into FP16 safetensors. It exists so you can run Bonsai in ordinary Hugging Face tooling or in frameworks that cannot yet read 1-bit weights. The FP16 files come to about 3.4 GB.

Who it suits: It suits developers who need Bonsai in a standard Hugging Face setup and cannot use the 1-bit builds yet.

Industry | Who uses it | How they use it | Example
Software | Developers | Load Bonsai in a framework that lacks 1-bit support | Test prompts against the model in their existing pipeline
Research | ML researchers | Work with the weights in plain FP16 form | Inspect the files with their usual Hugging Face tools
Tool building | Framework maintainers | Check how their tool handles the model before 1-bit kernels arrive | Run it through a standard loader

Anyone who can run the native 1-bit models should skip this copy, because the card recommends those instead.

Where to use it: You use it wherever Hugging Face safetensors models load.

• Standard Hugging Face tooling and frameworks without native 1-bit support
• Windows, macOS or Linux, with Python, for downloading the files
• For the recommended 1-bit builds: llama.cpp on CUDA, Metal or CPU, and MLX on Apple Silicon

Use case | Who | What they get back
Run Bonsai in standard tools | Developers | The model working at full FP16 size
Study the weights | Researchers | Plain safetensors files they can open anywhere
Bridge until 1-bit support lands | Tool maintainers | A stand-in that loads today

How to set it up: Setup is a download, nothing more.

• Python, for the Hugging Face hub tools
• Disk space for the FP16 files, about 3.4 GB

1. Install the Hugging Face hub tools.
2. Download the repository files.
3. Load the model with your usual Hugging Face tooling.

pip install -U huggingface_hub
hf download prism-ml/Bonsai-1.7B-unpacked

How to use it: The card shows no run command. You load the files as you would any Hugging Face safetensors model. For better results on memory and speed, the card points to two native 1-bit releases.

Tool | What it does
Bonsai-1.7B GGUF Q1_0 | A 1-bit GGUF file that runs in llama.cpp on CUDA, Metal or CPU
Bonsai-1.7B MLX 1-bit | A 1-bit MLX build for Apple Silicon

Example: You want to try Bonsai in a framework with no 1-bit support. You download this FP16 copy, load it like any Hugging Face model and run your prompts. It works at full size and has none of the memory or speed gains of the 1-bit version.

For newcomers and experts: The 1-bit version is the compact one. This repository is the same model expanded back to FP16 so ordinary tools can open it. If your tool reads GGUF or MLX, use the 1-bit build, and otherwise download this one.

For someone who knows the field, the 1-bit kernels live in the authors' forks of MLX and llama.cpp. The card says this unpacked copy becomes unnecessary once those land upstream. The card claims the 1-bit format cuts memory use by up to 14 times and runs about three times faster, with less energy per token. It says this FP16 copy has none of those gains.

1-bit Bonsai weights
Unpacked into FP16 safetensors
Loaded by standard Hugging Face tools
Runs at full size

Good to know: The record estimates about 860 MB for a 4-bit copy. The card names no language, task or benchmark.

• License: Apache-2.0
• Cost to run: none named
• Data it reaches: it downloads files from Hugging Face. The card names no account, web or clipboard access.
• Not for: anyone who can run the native 1-bit models, which the card recommends over this copy

On Hugging Face, 13 November 2026

• Parameters: 1.7B
• Weights: about 3.4 GB in F16; about 860 MB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 2,362
• Likes: 15
• License: apache-2.0
• Quantized versions: 13
• Fine-tunes: 4
• Tags: safetensors, qwen3, prismml, bonsai

Links

• Model: https://huggingface.co/prism-ml/Bonsai-1.7B-unpacked
• Files: https://huggingface.co/prism-ml/Bonsai-1.7B-unpacked/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download prism-ml/Bonsai-1.7B-unpacked

Then follow the usage in its model card. Nothing in this kit runs by itself.