larpers.larp
Studio lettersIssue 001 / 09.11.2026

491 MB

A language model small enough to take home.

01DownloadModel weights
02LoadBrowser memory
03GenerateOn your device
Fig. 1 / Signal path for a local reply

The second question

We asked our first browser build for the name of the Red Planet. It gave us one word: Mars. Then we asked for a two-line poem about that planet. It remembered Mars and wrote four lines.

Both results mattered. We could see that the conversation carried over. We could also count exactly where the instruction failed.

Inside the tab

The first send fetches roughly 491 megabytes of quantized Qwen weights. A WebAssembly worker reads them and generates the reply on the visitor’s own device. A second question can use the same weights. The transcript lives in the tab.

The working window holds 2,048 tokens. We reserve 384 for the answer and fit recent complete turns into the space that remains. A long conversation can still be on screen after its oldest lines have left the model’s context.

Try the same test

Give the model a short constraint you can check yourself. Ask for two colours, a six-word sentence, or a question without the letter E. Read the answer twice. Count once.

Open Larp Chat