AI Writer: Summarise, Rewrite and Fix Grammar

Writing help from an AI model that runs in your browser: summaries, a different tone, grammar fixes, emails, replies and social posts. The model downloads once; your text never leaves your device.

  • Downloads the AI model from Hugging Face once; your text never leaves this device
  • Free, no sign-up
  • Works on phones

Looking for an AI model on this device…

0 / 1,20,000 characters · 0 words
Open a text file
or drag and drop a .txt or .md file
Start an AI model above first.
Result
The result appears here.

What goes online

  • Downloads the AI model from Hugging Face once; your text never leaves this device.
  • Everything else happens inside this browser tab.

Often used next

Looking for something else?

DeskTray has hundreds of free tools that work the same way.

How to use the AI Writer

  1. Download an AI model once in Settings (the tool takes you there; the recommended one suits your device).
  2. Pick what you want to do, then type or paste your text, or open a text file.
  3. Press the button and watch the result appear. You can stop it at any time.
  4. Copy the result, or use it as the input for another step, such as fixing the grammar of a translation.

Features

  • Summaries in three lengths, including long text that is read in parts
  • Rewrite in six tones, fix grammar and spelling, shorten or expand
  • Emails, replies and WhatsApp, LinkedIn, Instagram or X posts from your points
  • Simple translation into 20 languages, with honest notes on quality
  • Streams as it writes, with Stop, Copy and “use result as input”
  • Works offline after the one-time model download

What the AI Writer can do

Pick a task, give it your text, and the model writes the result while you watch. You can stop it at any point and keep what it has written so far.

  • Summarise: a short summary in two to four sentences, bullet points, or a detailed version. Long text, such as a report or a chapter, is read in parts and the notes are combined into one summary.
  • Rewrite in a different tone: formal, friendly, concise, polite, persuasive or in simple words. Useful for turning a rough WhatsApp message into a note for your manager, or the other way round.
  • Fix grammar and spelling: corrects mistakes while changing as little as possible, so your wording stays yours.
  • Shorten or expand: halves a long message, or adds explanation to a short one without inventing new facts.
  • Write an email from points: type the points, choose the tone, and get a subject line, greeting and sign-off. Names it doesn’t know become placeholders like [Your name].
  • Write a reply: paste the message you received and, optionally, what you want to say.
  • Social posts and captions: a WhatsApp message, a LinkedIn post with hashtags, an Instagram caption or a short post for X.
  • Simple translation into 20 languages, including Hindi, Bengali, Marathi, Telugu, Tamil, Gujarati, Kannada, Malayalam, Punjabi, Odia and Urdu.

“Use result as input” chains steps: translate a message, then fix its grammar, or summarise a report and then rewrite the summary formally.

How it works, and why your text stays private

Most AI writing tools send what you type to a company’s servers. This one does not. The first time, you download a language model in Settings (the tool takes you there), and your browser fetches it from Hugging Face, the public library where AI models are published. The model is stored in your browser’s storage for this site, and from then on it runs on your own device, using the graphics chip through a browser feature called WebGPU.

Because the model is on your device, your text is never uploaded, there is no account or daily limit, and the tool keeps working without an internet connection once the model is downloaded. That makes it suitable for office emails, personal letters and anything you would rather not paste into an online chatbot.

Which model to choose

Settings marks the model that suits your device best, based on its memory, graphics chip and free storage. Qwen3, Ministral and Gemma 4 are open models under the Apache 2.0 licence; Gemma 3 and Gemma 2 come with Google’s Gemma Terms of Use. You can switch in Settings at any time, and every AI text tool on the site shares the model you downloaded.

  • Qwen3 0.6B (337 MB download, about 1.4 GB of memory while running): the fastest. Fine for short summaries and quick fixes on phones, but it makes more mistakes.
  • Gemma 3 1B (575 MB, about 0.7 GB): Google’s smallest Gemma 3. Quick and light on memory, for short answers and summaries in English; it needs a graphics chip with 16-bit support.
  • Qwen3 1.7B (940 MB, about 2 GB): a good balance for most laptops and recent phones.
  • Ministral 3 3B (1.8 GB, about 2.8 GB): clearer writing in English and European languages, but weak in Indian languages.
  • Qwen3 4B (2.1 GB, about 3.4 GB): the best quality here for its size and the best choice for Indian languages, for computers with 8 GB of memory or more.
  • Gemma 4 E2B (2.6 GB, about 3.4 GB): Google’s newest Gemma in its smallest size, with well-rounded answers in many languages.
  • Gemma 2 9B (4.9 GB, about 6.3 GB): the largest model here, with rich, careful writing in English. Slow, and only for computers with a strong graphics chip and plenty of memory.

Some graphics chips cannot use 16-bit numbers. On those, the tool automatically uses a 32-bit build of the same model, which needs a little more memory (Gemma 3 1B has no 32-bit build, so it isn’t offered there). A 940 MB download takes about three minutes on a 50 Mbps connection, and happens only once.

Honest limits

  • These models are small compared with online chatbots. They are good at reshaping text you give them and much weaker at knowing facts. They can get details wrong or invent them, so check names, numbers, dates and amounts before you send anything.
  • Indian languages are the weakest area. Expect awkward or wrong words, especially in Telugu, Tamil, Kannada, Malayalam and Odia. Use translations as a first draft and have important text checked by a fluent speaker.
  • The model reads about 3,000 English words at a time. Longer text is split into parts: summaries then lose some detail, and rewrites are done part by part.
  • Length limits keep waiting times sensible: up to 1,20,000 characters for a summary, 15,000 for rewriting, grammar, shortening, expanding and translation, and 4,000 for emails, replies and posts.
  • Speed depends on your graphics chip. The first start after opening the page takes a few seconds while the model loads into memory; larger models write more slowly.

What your device needs

WebGPU is available in Chrome and Edge on Windows, Mac and Chromebooks, in Chrome on recent Android phones, and in Safari on iPhone, iPad and Mac with iOS, iPadOS or macOS 26 or later. On Linux it may need to be switched on in the browser. If the tool says WebGPU is missing, updating the browser often helps.

You also need free storage for the model and enough memory to run it. A laptop with 8 GB of memory runs every model; phones with 4 GB are best with the smallest. Settings checks this for you and explains when a model is too big.

Tips for better results

  • Put details in “Anything else?”, for example “mention the meeting is on Monday at 11” or “keep it under 100 words”.
  • For emails, write one point per line. Short, clear points give the best drafts.
  • If a summary misses something, ask for the detailed version or summarise a smaller part.
  • If the result is cut short, split the text and run each part separately.

Frequently asked questions

Is my text uploaded anywhere?

No. The AI model runs inside your browser on your own device. Only the model itself is downloaded from Hugging Face, once. What you type is processed locally and never sent to a server, and the tool works offline after the download.

Which model should I choose?

The one marked Recommended fits your device’s memory, graphics chip and free storage. Qwen3 1.7B (940 MB) is a good balance on most laptops and recent phones; Qwen3 4B (2.1 GB) writes best on computers with 8 GB of memory or more; Gemma 2 9B (4.9 GB), the largest, needs a strong graphics chip; Qwen3 0.6B (337 MB) is fastest on older phones.

Why does it say my browser has no WebGPU?

The models run on the graphics chip through WebGPU. It is available in Chrome and Edge on Windows, Mac and Chromebooks, in Chrome on recent Android phones, and in Safari on iOS, iPadOS and macOS 26 or later. Older browsers cannot run them.

Is this the same as ChatGPT?

No. It is a much smaller model that runs on your own device, so it knows less and makes more mistakes, but your text stays private, it is free without limits and it works offline after the download.

Can I use it without the internet?

Yes. You need a connection once, to download the model. After that, the AI Writer works offline in the same browser.

How do I free the space the model uses?

Open Settings (the gear at the top of every page). It lists every downloaded model with its size and a Delete button, and shows how much storage DeskTray uses in this browser.

More AI tools

All 4 AI Tools