AI Text Summarizer — Free, No Signup, Sources Cited

The page runs two complementary summarization methods on the same input and shows both. They are good at different things, and reading them side by side is the fastest way to judge whether the summary can be trusted on a particular passage.

1. DistilBART abstractive (server-side)

The TL;DR paragraph comes from sshleifer/distilbart-cnn-12-6, a 12-layer-encoder / 6-layer-decoder distillation of BART fine-tuned on the CNN/DailyMail summarization corpus (Shleifer & Rush, 2020). We call the Hugging Face Inference API (api-inference.huggingface.co) from a server-only Next.js route handler with do_sample=false for deterministic output. The browser never downloads model weights — only your text is sent, and only a small JSON response comes back. Generation length is bound by the length preset you pick:

• Short: min 40 tokens, max 80 tokens, 3 extractive bullets.
• Medium: min 70 tokens, max 140 tokens, 4 extractive bullets.
• Long: min 110 tokens, max 220 tokens, 5 extractive bullets.

2. TextRank extractive (always on)

The bullets come from TextRank (Mihalcea & Tarau, EMNLP 2004) — a PageRank-style algorithm run over a sentence-similarity graph. Text is split into sentences on terminal punctuation. Each sentence becomes a node; the weight of the edge between two sentences is the Mihalcea & Tarau similarity:

               |S_i ∩ S_j|
sim(S_i, S_j) = ─────────────────────────
                 log|S_i| + log|S_j|

where S_i is the set of content tokens in sentence i (after Unicode normalisation, lowercasing, and removing the 126-word Snowball English stop-word list). PageRank then iterates with damping factor 0.85 until the scores stop changing (tolerance 1e-4 or 30 iterations, whichever comes first). The top-scoring sentences are returned in original document order — that keeps the bullets reading naturally instead of being shuffled by importance.

TextRank is pure JavaScript with no API call, so it always runs even when neural inference is not configured on this build. It is also fully deterministic — the same text always produces the same bullets — which is useful when you need a summary you can reproduce offline or when comparing two versions of a document.

3. Non-English handling

DistilBART-CNN was trained on English news and the stop-word list is English-only, so the page is honest about its limits. Sentences whose letters are more than 70% non-Latin script are flagged in the breakdown and excluded from the bullet ranking. For Sinhala or Tamil text we recommend translating to English first with our AI Translator and then summarizing. The TL;DR still runs through the model on mixed content — output quality just drops.

4. Compression and reading-time stats

Alongside the summary the page reports four numbers so you can judge the result at a glance. Key points is the bullet count over the total sentence count. Compression is 1 − (summary words ÷ original words) — a 1,000-word article cut to a 120-word summary reads as 88% compression. Read time saved converts both word counts to minutes at 220 words per minute, the same reading speed our reading-time estimator uses, then shows the difference. Word counts come from the same Unicode-aware token regex that powers the word counter, so the figures line up if you check them independently.

The summarizer is one half of a workflow. To pull the standalone terms a passage keeps returning to, run the bullets through the AI keyword extractor; to check that a draft summary fits inside a model’s context window before you paste it into a chatbot, the AI token counter gives the exact token count. DistilBART itself reads at most 1,024 tokens, which is why very long inputs are truncated rather than rejected — see the worked example below.

Both passes are deterministic given the same input and length preset. The HF Inference API is stateless: every call is independent and no identifiers are sent. If the API is rate-limited or unreachable, the extractive bullets and the top-ranked sentence as a TL;DR fallback are still returned so the page never shows a dead state.

• Hugging Face — sshleifer/distilbart-cnn-12-6 (model card)
• Shleifer & Rush, 2020 — Pre-trained Summarization Distillation (arXiv:2010.13002)
• Mihalcea & Tarau, 2004 — TextRank: Bringing Order Into Texts (EMNLP)
• Hugging Face Inference API — official documentation
• Snowball English stop-word list (public domain)
• CNN/DailyMail summarization dataset — See et al., 2017

The model card, API endpoint, and references were last cross-checked on 2026-05-12. The page is reviewed whenever the Hugging Face Inference API contract changes or the upstream model repository is re-uploaded.