GlowSpark
A tiny AI we built from scratch. It writes fairy tales right here on your device: no account, no server, no waiting.
Start a tale
Your tale
Tap any word to see what else the model considered, and send the tale down that path. Dotted words are the ones it was least sure about.
Instead, the model could have written
The numbers, measured
Nothing here is an estimate. Every figure is read from the benchmark file our scripts wrote, and the two marked with a green dot are timed on your device as you use it.
How well does it predict stories it never read?
Where it wins, where it loses
GlowSpark is not a rival to ChatGPT, Claude or Gemini at knowing things. It goes after what they cannot do: start instantly, run anywhere, work offline and keep every word on your device.
| GlowSpark | Cloud chatbotsChatGPT, Claude, Gemini | Big models in the browserWebLLM | Chrome's built-in AIGemini Nano | |
|---|---|---|---|---|
| Knowledge and reasoning | None. It has only read old fairy tales. | The best there is | Good, for small models | Good, for a small model |
| Before the first word | One small download | Sign-in for most, then a round trip to a data centre for every reply | A model download that is many times larger | A model download managed by Chrome |
| Runs on | Any modern browser, phones included. Plain JavaScript on the CPU. | Anything with a connection | Browsers and devices with WebGPU | Desktop Chrome on supported hardware |
| Works offline | Yes, after the first visit | No | Yes, after the download | Yes, after the download |
| Where your words go | Nowhere. There is no server to send them to. | To the provider's servers | They stay on the device | They stay on the device |
| Account, key or usage limit | None | Usually an account, and limits on free plans | None | None |
| Can you see why it chose a word? | Yes. Tap any word for the alternatives and their odds. | Not in the chat apps | Only if the app builds it | No |
Sources for the other columns, checked on …: WebLLM README (read directly), Chrome built-in AI requirements (through a web search, because the page would not open from the network this was built on). We did not benchmark those systems ourselves, so no speed, size or quality number is claimed for them.
What it can't do
- It cannot answer questions. It knows no facts, only how old tales tend to go. For answers, use GlowChat.
- It writes English only, in the voice of books from a century or more ago, and it carries their old-fashioned ideas too.
- It loses the thread fast. The phrases sound like a fairy tale, but sentences often go wrong, and characters and plot change from line to line.
- It remembers only the last few hundred words, so a long tale forgets how it began.
Built from scratch
- The tokenizer: our own byte-pair encoder, trained on the books. No tokenizer library.
- The model: a small transformer that started from random numbers. No pretrained weights, nothing downloaded from another model.
- The training: one laptop processor, no graphics card, public-domain books only.
- The engine: plain JavaScript written for this page. No AI library, no WebGPU, no WebAssembly.
The books it learned from
All are in the public domain in the USA and come from Project Gutenberg. Books marked “held out” were never shown to the model: they are the test.