AI Grammar Checker & Paraphraser

About the AI Grammar Corrector

This AI Grammar Corrector fixes grammar entirely on your device. It loads the Xenova/t5-base-grammar-correction model with Transformers.js and runs it inside a Web Worker using WebGPU when available, or WebAssembly otherwise. Your text is never uploaded to any server: there is no API call, no logging, and no telemetry on your input.

The tool shows a word-level diff between your original and the corrected text, a change-rate metric, and the per-run inference time and active backend. You can copy the corrected text or download a structured JSON or plain-text report for your own records. See also our AI Text Summarizer and AI Translator, which use the same in-browser stack.

Is my text uploaded to a server when I use this grammar corrector?

No. The tool runs entirely in your browser using Transformers.js and WebAssembly (or WebGPU). The model is downloaded once from the Hugging Face CDN and cached locally by your browser, after which all correction happens on your CPU or GPU without your text ever leaving the device. There is no backend API call, no logging, and no telemetry on your input. This matters for legal contracts, medical notes, internal memos, and student essays. You can verify it yourself by opening DevTools, watching the Network tab, and confirming that no request is sent containing your text after the model finishes loading.

Which model does the tool use under the hood?

It uses Xenova/t5-base-grammar-correction, a T5-base encoder-decoder model (about 220M parameters) fine-tuned for grammatical-error correction and served through Transformers.js. On WebAssembly it runs in q8 (8-bit) quantization to keep the download and memory footprint small; on WebGPU it runs in fp32. Each input is prefixed with the model's expected "grammar: " task token. Longer text is split by sentence so every pass stays under the 512-token context window.

How accurate are the corrections, and what are the limits?

The model reliably fixes common errors: subject-verb agreement, articles, prepositions, verb tense, duplicated words, and basic punctuation. It is English-centric and was trained on English grammar-correction data, so results on other languages are unreliable. As a base-size model it can miss subtle style issues, idiomatic phrasing, and domain-specific jargon, and it may occasionally rephrase rather than minimally edit. Treat the output as a fast first pass and proofread before publishing critical documents. These are general expectations, not guaranteed benchmark scores.

What is the word-level diff and the Change Rate?

After correction, the tool aligns your original and corrected text token by token using a longest-common-subsequence diff. Removed words are shown struck through in red and added or changed words in green, so you can see exactly what the model touched. Change Rate is the number of changed tokens divided by the number of words in your original text, expressed as a percentage. A low change rate means the model made only light edits; a high rate means it rewrote more of your text, which is worth reviewing closely.

Can I run the model on my GPU for faster inference?

Yes. Transformers.js detects WebGPU support (Chrome 113+, Edge, and other Chromium browsers) and uses it automatically; otherwise it falls back to WebAssembly. The active backend is shown as a badge next to the input (WebGPU or WASM), and each run reports its inference time in milliseconds. WebGPU is typically several times faster than WASM on the same hardware, especially for longer paragraphs. No configuration is required: the tool picks the fastest available backend at run time.

What is the maximum input length, and does the tool work offline?

There is no hard character cap, but the model has a 512-token context window, so the tool splits long input into sentence-sized chunks and corrects each one in turn; very long documents simply take more passes. Once the model has loaded the first time, its files are cached by your browser, so you can disconnect from the internet, reload the page, and keep correcting offline for as long as the cache survives. This makes it useful on planes, on secure networks, or with intermittent connectivity.

Can I export the corrections for my workflow?

Yes. After a correction you can copy the corrected text, or download a report as JSON or plain text. The JSON report contains the model name, the active backend, the inference time, your original and corrected text, and the structured list of added and removed tokens, so editors and reviewers can audit exactly what changed and feed it into a track-changes or review workflow. The plain-text report is a human-readable summary of the same information.