QuantBench

QuantBench

A browser studio for exploring LLM quantization below one byte per weight, with a Pareto frontier explorer, BitNet 1.58b bit-packing registers and a layer sensitivity simulator. For engineers and students.

Made by prajwalpujari16 on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: QuantBench is a browser studio for exploring LLM quantization below one byte per weight. It offers analytical exploration before you quantize anything, bit-packing registers for BitNet 1.58b ternary weights, and a simulator for how sensitive each layer is. It is a static page, so nothing runs on a server.

Who it suits: It suits people who want to see what happens to a model's quality and size when its weights use fewer bits.

Title: Who might use this studio
Industry | Who uses it | How they use it | Example
ML engineering | Engineers | Weigh compression against quality before quantizing | Explore the frontier of size against signal quality
Education | Students | Learn how low-bit weights are packed | Watch ternary values fill a register
AI research | Researchers | Reason about which layers tolerate compression | Try the layer sensitivity simulator

If you need to quantize a real model file, you can skip it, since this is an analytical explorer.

Where to use it: You can use it online or open its files on your own computer.

• Web browser: the Space's page on Hugging Face, with nothing to install
• Windows, macOS or Linux: the downloaded files, opened in a browser
• Python: a PyPI package of the same name, named in the readme

Title: What the studio is good for
Use case | Who | What they get back
Find a compression trade-off | Engineers | A frontier that plots compression against SQNR
Learn bit packing | Students | Live registers for BitNet 1.58b, INT4 and 2:4 sparsity
Spot fragile layers | Researchers | A simulation of per-layer sensitivity

How to set it up: The hosted page needs nothing. To get the files you need Python and pip.

• Python, with pip
• Any web browser

1. Install the huggingface_hub package.
2. Download the Space's files into a folder.
3. Move into that folder.
4. Open index.html in a web browser.

pip install -U huggingface_hub
hf download prajwalpujari16/quantbench --repo-type space --local-dir quantbench
cd quantbench

The readme also names a PyPI package, which installs with this command.

pip install quantbench

How to use it: Open the Space online, or open index.html from your downloaded folder. Then try the three interactive tools, starting with the frontier explorer.

Title: The interactive tools
Tool | What it does
SQNR against compression Pareto frontier explorer | Shows the trade-off between signal quality and compression
Bit-packing registers | Show BitNet 1.58b, INT4 and 2:4 sparsity packing live
Transformer layer sensitivity simulator | Simulates how sensitive each layer is to quantization

Example: You open the Pareto frontier explorer and look at how signal quality drops as compression rises. Then you open the bit-packing registers to see how BitNet 1.58b ternary values are packed together.

For newcomers and experts: If you are new to this, remember that quantization stores weights in fewer bits to save memory. Start with the frontier explorer and move along it slowly.

If you know the field, the page pairs an SQNR against compression frontier with live packing registers and a per-layer sensitivity simulator. It names Qwen2.5 0.5B Instruct and Llama 3.1 8B, and a Pareto dataset of its own.

Open the studio
Pick a tool to explore
Adjust bits, layers or settings
Read the trade-off it shows

Good to know: The exploration is analytical, so it does not touch a real model file.

• Static page: plain HTML and CSS, with no server code
• Models: the Space names Qwen2.5 0.5B Instruct and Llama 3.1 8B
• License: apache-2
• Cost to run: none named
• Data it reaches: nothing beyond its own files is named in the notes
• Not for: quantizing a real model file

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: apache-2.0

Links

• Space: https://huggingface.co/spaces/prajwalpujari16/quantbench
• Files: https://huggingface.co/spaces/prajwalpujari16/quantbench/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.