AI Text Summarizer
Summarize a long article, an email thread, or a report down to a few sentences. Pick short, medium, or long; the AI model downloads once and then works offline.
Runs entirely in your browser — your files and text never leave your device. The AI model downloads the first time you run this tool, then stays on this device — so later runs work offline.
Downloads the summarising and language models (~1251 MB) the first time you run it, then works offline.
Runs on your device — the model downloads once on first use, then works offline. Standard is a distilled BART model that works everywhere. Sharp is a language model that reads the whole passage and writes a more faithful summary, but it needs a browser that can reach your graphics card; the picker greys it out when it can't. Sharp trims long text at about 1,000 words; Standard reads it in overlapping pieces.
How to use
- Drop the article, thread, or report into Text to summarize. The box opens nine rows tall and you can drag the corner to make it bigger.
- Set Summary length, then look at the Model row. Sharp is preselected wherever it can run, and greyed out with the reason written next to it where it cannot.
- Press Summarize. Compare what lands in the Summary box against the original, then use the copy button to its right.
About this tool
Some documents you only need the gist of. A forty-paragraph release note. A mail thread you got added to at five on a Friday. A paper you have to cite but will never read closely. Paste it into the text summarizer, choose a length, press Summarize. Short comes back as a sentence or two, medium as three or four, long as six to eight. With the Sharp model you watch the summary get written a few words at a time, and there is a Cancel button beside the progress bar if it starts wandering.
Two models sit behind the Model picker, and the picker prints what each one costs to fetch. Standard is roughly 708 MB, Sharp roughly 543 MB. Standard is a distilled BART, an older summarizing model that loads on any browser including Safari and phones. Sharp is Qwen3-0.6B, a small language model that reads the passage and then writes about it, which is why it holds together over longer text. Sharp wants WebGPU, so the page selects it for you when your hardware can reach the graphics card and drops back to Standard when it cannot. Whichever one you use downloads a single time, then sits in the browser cache and works with the network off.
A 2,000-word release note, run through both models
Paste a release note of about 12,500 characters, close to 2,000 words, and leave the length on Medium. With Standard selected the worker cuts the text into four pieces of 3,600 characters that overlap by 400 so nothing is lost at a seam, and asks the model for 33 to 70 tokens on each. Those four partial summaries are joined and summarized once more inside the medium window of 55 to 100 tokens, five model passes in all, and the result is trimmed to its first four complete sentences before it lands in the Summary box.
Switch to Sharp and the shape changes. The paste is cut at word 1,000, so the second half of the note is never read. The model is asked for three or four sentences under a 224-token ceiling, and the summary appears in the box a few words at a time while it writes. The first run on either model begins with Downloading model and a percentage; once the files are cached that stage is over almost before you see it and the status moves on to Summarizing. Press Cancel partway and the status line reads Canceled, with whatever text had streamed left in the box for you to keep or clear.
Good to know
- Sharp is handed the first 1,000 words of your paste and nothing after that. Split a long report at its section headings and run the parts separately.
- Standard's long-text route squeezes each 3,600-character piece down to 50, 70 or 126 tokens, depending on the length you chose, before it has seen the whole. A fact that appears once, deep in a long paste, can be lost at that first stage and never reach the final summary.
- Standard is trimmed by sentence count after the fact: two sentences for short, four for medium, everything it wrote for long. A final sentence the model left dangling is dropped rather than shown half-built.
- Standard is a 2019-era summarizer. On dense or technical prose it will flatten a qualifier, or state something the source never quite said.
- Sharp is 0.6 billion parameters. It stays much closer to the source than Standard does, and it is still small enough to misread a contract clause or a methods section.
- That first download hurts on a metered connection. Standard is the bigger of the two, which catches people out, since the better model is the smaller file.
- Output is one block of prose. There is no bullet mode, and no way to trace a claim back to the paragraph it came from.
Things people ask
Should I pick Standard or Sharp?
Take Sharp any time the picker lets you. It reads the whole passage and writes about it, so it stays coherent on long input where Standard starts repeating the opening paragraph. Standard is there for machines and browsers with no usable GPU, which mostly means older phones and older Safari.
How much text can I paste at once?
The two models answer differently. Sharp cuts your text at word 1,000 before it starts, so the tail of a long article is never read at all; summarize section by section for anything longer and join the pieces yourself. Standard has no cap. Anything past 3,600 characters is split into overlapping pieces, each summarized on its own and the partials summarized again, so a 5,000-word report goes through whole at the cost of one model pass per piece plus one for the final round.
Why is the model download so large?
The model is the download. Standard weighs about 708 MB and Sharp about 543 MB, and the Model picker prints those figures beside each name so you know before you commit. It happens once per browser. After that the file is cached and a summary starts in seconds.
The summary contains something the article never said. What happened?
Compact summarizers work by predicting likely wording, and a prediction can land on a claim that was never in the source. Standard does it more often than Sharp, especially on technical passages. Read the summary next to the original before you forward or cite it.
What do short, medium, and long actually change?
Both the instruction and the budget. Sharp is asked for one or two sentences, three or four, or six to eight, with a token ceiling of 96, 224, or 448 to match. Standard is given a token window instead, 15 to 45 for short, 55 to 100 for medium, 110 to 180 for long, and then a sentence trim. Real length still moves with the input, because the model picks its own stopping points.
Will the same text give the same summary twice?
Both models decode greedily with sampling switched off, so identical text with the same length and the same model produces the same summary on every run. Change a word and the wording may shift further than you expect. Standard also refuses to repeat any run of three words, which is why its output never loops a phrase and occasionally why it bends a proper name that legitimately needs the same three words twice.
Why is Sharp greyed out on my machine?
The picker writes the reason after the name. Needs a browser with WebGPU support means the browser cannot hand the model to your graphics card at all. Needs a GPU with 16-bit shader support means WebGPU is there but the adapter lacks the half-precision shaders the Sharp weights are packed for. In either case Standard stays selected and does the job with the older model.
Missing something in AI Text Summarizer? Suggest a feature →