AI Summarizer
Paste a long article, get a tight summary. Runs the distilled BART model entirely in your browser — nothing uploaded, no API key.
Why a Summarizer That Runs in Your Browser
Server-based summarizers see every paragraph you paste. For an internal memo, a draft contract, a leaked transcript, or anything else you would not paste into a public chatbot, that is a problem worth designing around. This tool loads a distilled BART model — about 155 MB, downloaded once and cached — and runs every summary directly in your browser. The text you paste never travels anywhere. No request log. No retention policy. No "we may use your prompts to improve our models." If your browser supports WebGPU the inference uses your graphics card, which on a modern laptop is fast enough that a 2,000-word article summarizes in a handful of seconds. WebGPU-less browsers fall back to WebAssembly and run perfectly well, just a little slower. The model is distilbart-cnn-6-6 from Sam Shleifer at Hugging Face, trained on the CNN/DailyMail summarization dataset and released under the Apache 2.0 license. Apache 2.0 means commercial use is allowed without strings.
How the Summarizer Works
Click Load model the first time you use the tool. The browser downloads the model files — about 155 MB total, served from the Hugging Face CDN — and stores them in IndexedDB. Subsequent visits skip the download and load the cached files, which takes a few seconds rather than a few minutes. Paste your article in the text area. The model handles up to roughly 1,024 input tokens per pass (about 800 words of typical English prose); longer inputs are automatically chunked, each chunk summarized, and the chunk summaries concatenated. Two length controls let you bias the output toward shorter or longer summaries — minimum and maximum new tokens — and a beam-search toggle trades speed for slightly more polished phrasing. Output appears in the result panel with a copy button. The model was trained on news articles, so it does best on factual, structured prose. Marketing copy, transcript fragments, and code-heavy text are out-of-distribution and produce weaker summaries — for those, paraphrasing or grammar-correcting a draft you wrote yourself is usually a better fit.
Frequently Asked Questions
How big is the model download?+
The summarizer model is approximately 155 MB, served from the Hugging Face CDN. The download happens once on first use — after that the model is cached in your browser and loads in a few seconds for subsequent sessions.
Is the article I paste in sent anywhere for summarization?+
No. After the model files finish downloading on first use, every summary runs entirely in your browser. The text you paste never leaves your machine and is not sent to any server, including our own.
What model and license does the summarizer use?+
The summarizer uses distilbart-cnn-6-6 from Sam Shleifer at Hugging Face, a distilled version of BART fine-tuned on the CNN/DailyMail dataset. It is released under the Apache 2.0 license, which permits commercial use.
What input length does the summarizer accept?+
The model handles about 1,024 tokens per pass — roughly 800 words of typical English prose. For longer inputs the tool chunks the text, summarizes each chunk, and concatenates the chunk summaries. Very long inputs may take a minute or two on slower devices.
Why is the first summary slow but later ones fast?+
The first run includes the model download (about 155 MB) and a warm-up pass. Subsequent runs reuse the cached model and warmed-up runtime, so they only spend time on inference. On a modern laptop with WebGPU a cold start can take 30-60 seconds and a warm summary runs in 3-8 seconds.
Does this work on phones?+
Yes on iPhones running iOS 17+ and on modern Android phones, though performance is slower than on a laptop. WebGPU support on mobile is still uneven — Safari on iOS uses WebGPU on iPhone 15 Pro and later, and Chrome on Android uses it on most flagships from 2023 onward. WebAssembly fallback works everywhere else.
Can I summarize PDFs or Word documents directly?+
Not directly — paste the extracted text into the input area. For PDFs use a PDF-to-text tool first; for Word documents, copy-paste from the document. Adding native file parsing is on the v34 roadmap.
How do the summaries compare to ChatGPT or Claude?+
The hosted models from OpenAI and Anthropic produce noticeably better summaries — they are 10-100x larger and trained on much more diverse data. The trade-off is that this tool sends nothing to a server, requires no API key, has no rate limits, and works offline after the first model download. For sensitive content or high-volume use the privacy and cost advantages often outweigh the quality gap.
Built by Derek Giordano · Part of Ultimate Design Tools
Privacy Policy · Terms of Service