Next-Concept Modeling WebGPU

Next-Concept Modeling WebGPU

Runs small Next-Concept Modeling checkpoints from Hazy Research in your browser on WebGPU, on TinyStories, LM1B small and Sudoku. It suits researchers and developers who want a local, no install demo.

Made by mrfakename 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: Next-Concept Modeling WebGPU is a static web page that runs small Next-Concept Modeling checkpoints from Hazy Research. It handles TinyStories, LM1B small and Sudoku. All of it runs on your side, using ONNX Runtime Web on WebGPU.

Who it suits: People who want to try a coarse-to-fine text generator in a browser tab, without a server or a Python setup.

Industry | Who uses it | How they use it | Example
Research | ML researchers | Try the small checkpoints without installing anything | The Sudoku task
Software | Web developers | Study how the model runs in a page with a worker | Reading ncm.js and worker.js
Education | Students of language models | Watch a small generator work on a laptop | A TinyStories run
Outreach | Teachers | Demonstrate in-browser machine learning | A class demo

If you need the larger models, you can skip it, because a sister Space covers those.

Where to use it:

• A web browser with WebGPU, on the Space's page at Hugging Face
• Windows, macOS or Linux, opening index.html from a downloaded copy

Use case | Who | What they get back
TinyStories | Curious users | Short generated text from a small checkpoint
LM1B small | Researchers | Output from the small LM1B model
Sudoku | Hobbyists | A puzzle solved by the model in the page

How to set it up:

• A browser that supports WebGPU
• Python only if you want to download the files to your computer
• A connection for the first load, when the converted weights are fetched

1. Install the Hugging Face hub package.
2. Download the Space with the hf command.
3. Move into the downloaded folder.
4. Open index.html in a web browser.

pip install -U huggingface_hub
hf download mrfakename/ncm-webgpu --repo-type space --local-dir ncm-webgpu
cd ncm-webgpu

How to use it: Open the Space page, or index.html from your copy, and choose one of the three tasks. For the larger models, such as LM1B medium and large, XSum and WMT14, use the mrfakename/ncm Space instead.

Tool | What it does
index.html | The page you open in a browser
app.js and ncm.js | Script files for the page and the model logic
worker.js | A script file that can run work off the main page thread
sudoku_ex.json | Example data for the Sudoku task

Example: You open the page, pick the Sudoku task and run it. The small checkpoint works on your graphics card and the result appears in the page.

For newcomers and experts: Open the page in a WebGPU browser and try TinyStories first. Nothing needs installing, though the weights are fetched on the first visit.

The weights in mrfakename/ncm-webgpu-weights were converted from hazyresearch/ncm-checkpoints, which carry an MIT license. Inference runs client side through ONNX Runtime Web. The trade-off is that only the three small tasks fit the browser setup, while the bigger models live in the other Space.

Browser opens the static page
Converted weights load from a model repository
ONNX Runtime Web runs the model on WebGPU
Generated output shows in the page

Good to know:

• Covers only TinyStories, LM1B small and Sudoku
• Built as static HTML, so there is nothing to install to use it online
• License: mit
• Cost to run: none named
• Data it reaches: converted model weights from Hugging Face on first load, nothing else named
• Not for: the larger models, which the other Space handles

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: static
• License: mit

Links

• Space: https://huggingface.co/spaces/mrfakename/ncm-webgpu
• Files: https://huggingface.co/spaces/mrfakename/ncm-webgpu/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.