How a 270M-parameter claim analyzer runs entirely in your browser

What is in the 253 MB download?

The quantised weights of the fine-tuned model: a 270-million-parameter model stored as a 4-bit (Q4_K_M) GGUF file hosted on Hugging Face. The page shows a progress bar while it downloads, asks the browser to cache it, and remembers in localStorage that the model was cached, so a returning visitor's model loads without a click. In a test on 3 October 2026 (desktop Chrome, headless, empty cache) the model downloaded and loaded in 40.1 seconds; your time depends on your connection.

How does a browser run a language model?

wllama compiles llama.cpp to WebAssembly. The page loads its 8,457,512-byte wllama.wasm from this site and creates the model with a 2,048-token context. A claim is sent as a streaming completion with temperature: 0 and at most 512 new tokens, and the reply is rendered as it arrives. In the test the page was not cross-origin isolated, so multi-threaded WebAssembly was unavailable and generation ran on a single thread; the page's status line says "WebGPU if available" when the browser exposes navigator.gpu. Timings from your own hardware will differ.

What prompt does the model get?

The fine-tuned model has absorbed the task, so the page sends only a short header around the claim, the compact format from the model card:

Analyze this claim.

INPUT:
<your claim>

ANALYSIS:

Decoding is greedy and stops at the end-of-sequence token, so the same claim should give the same analysis each time.

How does it refuse non-English or off-topic input?

In two layers. First, a small script in the page checks the text before the model sees it: it refuses when more than a quarter of the letters are non-Latin, or, for texts of six words or more, when there are at least four accented letters and at most one common English word, or when a list of common German, French, Spanish, Italian, Portuguese, Dutch or Turkish words matches at least twice and more than one time more often than common English words do. The reply is "I only analyze claims written in English." Second, the model was trained with refusal examples: for "Write me a poem about the sea." it answered "I only analyze English-language claims. Please provide a claim or argument to analyze." in about four seconds.

How long does an analysis take?

One run each, single-threaded, on one machine. See the example output page for the text these produced.

What does the author report about quality?

The tool page reports a model trained on about 1,400 synthetic examples, including refusals for non-English and off-topic input, and a section-header F1 against reference answers that rose from 0.42 to 0.91 on 68 held-out prompts. That score measures whether the output has the right section headings, not whether the content is correct. The page also states plainly that the model is small and wrong sometimes, and the example output page shows a real mistake.