README.md

---
title: GPT-2
emoji: ✍️
colorFrom: gray
colorTo: indigo
sdk: static
app_file: index.html
pinned: false
short_description: GPT-2 completion + next-token probs, in-browser
---

# GPT-2

A demo of [openai-community/gpt2](https://huggingface.co/openai-community/gpt2) — the
original 124M-parameter GPT-2 — running **entirely in your browser** via
[transformers.js](https://github.com/huggingface/transformers.js). No server, no API key,
nothing leaves your machine.

Two views of the same model:

- **Completion** — streams a continuation of your prompt token by token, with
  temperature / top-p / top-k / repetition-penalty controls.
- **Next token** — the raw probability distribution GPT-2 assigns to the next token.
  This is all the model actually computes; generation is just doing it in a loop.

## Implementation notes

Weights come from [`Xenova/gpt2`](https://huggingface.co/Xenova/gpt2), the ONNX export of
the same checkpoint (the official repo's ONNX files are fp32-only at 655 MB, too heavy for
a browser download). The app picks its runtime at load time:

| Browser | Backend | Precision | Download |
|---|---|---|---|
| WebGPU available | `webgpu` | fp16 | 251 MB |
| otherwise | `wasm` | int8 | 281 MB |

Cached by the browser after the first run. Prompts are truncated to keep
prompt + generation inside GPT-2's 1024-token context window.

## Also in this repo

`gradio-zerogpu/` holds an equivalent **Gradio + ZeroGPU** implementation (server-side
PyTorch, streaming, and an MCP server). It isn't what this Space runs — hosting a Gradio
Space requires PRO or an account older than 30 days — but it's ready to deploy as-is:

```bash
hf repos create <ns>/gpt2-gpu --type space --space-sdk gradio --flavor zero-a10g --public
hf upload <ns>/gpt2-gpu ./gradio-zerogpu . --repo-type space
```