Including the ones with awkward answers, what is still a placeholder, what the numbers cannot tell you, and why some of them are deliberately not headlines.
A thesaurus for the age of language models. Look up a word and see how AI models actually split it into tokens, which words they place near it, where it sits on interpretable axes, and how its usage compares between people and machines. A friendly Story view by default, with every raw number one tap away in the Expert view.
Writers and editors who want to know which words read as machine-written; teachers and students; and NLP people who want the cosine tables, the z-scores and the API.
20,660: the most frequent words that have a WordNet entry, plus every scientific term WordNet's own topic domains supply, plus a curated set kept so every panel has a worked example.
Qwen3 8B (Alibaba, CN), DeepSeek V3 (DeepSeek, CN), Mistral 7B (Mistral, EU). Three labs, two regions, so "models disagree" is something you can actually see rather than a claim.
Each model's real model.embed_tokens matrix, extracted from a
single weight shard. That needs no GPU and no inference, which is why a
200-billion-parameter model costs the same to analyse here as a small one.
Because they are, to a language model. hot and cold appear in almost identical sentences, so they sit close together in embedding space. Cosine similarity cannot separate them, which is exactly why opposition here is seeded from a dictionary and only ranked by geometry.
Probably because that model splits the word into pieces. Embeddings are per
token, so a word the tokenizer breaks up has no vector of its own and
we average its fragments. It shows: Mistral splits tapestry into
tap·est·ry and its nearest neighbour becomes incest, a
shared est fragment, not a shared meaning. Any panel built on a
split word says so. The useful lesson is real: the models that split a
word are the ones that represent it worst.
More formal than 90% of the vocabulary here. Raw projections onto an axis are tiny and clustered, so taken literally every word sits "in the middle" of everything. The scores are percentile ranks instead, which also makes them comparable between models whose embeddings have different scales.
Full detail on the methodology page.
No. Every word page is free, unlimited and needs no login: that will not change. Accounts and keys exist for the API.
Anonymous callers get 60 API requests a day. A free account gets 500 a day. A 72-hour demo key is one click away once your email is confirmed and gives 1,000 requests. Paid plans raise that and unlock the endpoints a static site cannot serve: text scoring, bulk lookup and CSV export.
Very little, and all of it is listed in the privacy policy below: an email address and a password hash if you have an account, key hashes, and a salted daily-rotating hash of your IP address for rate limiting.
Yes, and please cite the methodology page alongside it, including which parts were measured on the date you looked. Every extraction is dated, so results stay comparable over time.
Spanish, French, Japanese and Chinese, on top of English. Look up perro or 犬 and you get a page, but it is worth being exact about what is and is not measured on your word.
The English word is the pivot. Every language is joined to English WordNet through the shared sense, so perro resolves to the sense dog carries, and the neighbours, personality and usage numbers you see are the ones measured on dog. The page says so at the top, every time. There is no separate Spanish embedding behind it.
One panel is genuinely yours: the token split. A tokenizer does not care what language it is fed, so that panel splits your headword, not the English one — and it is where the languages differ most. dog is one token; perro and chien are two; 犬 is three raw bytes to Qwen and Mistral and a single token to DeepSeek. Across the whole build, 98% of foreign headwords split into more than one piece, against 66% of English words on Mistral.
Definitions stay English. Open Multilingual WordNet supplies headwords, not glosses, so a Spanish page shows a Spanish word above an English definition. Per-language definitions are a different and much larger data job.
Coverage is partial and uneven, measured against the 20,660-word vocabulary: Japanese 92.0%, French 91.8%, Spanish 74.9%, Chinese 71.4%. A word with no headword in a language shows that language as a dashed gap rather than guessing.
Yes. LLMsaurus is for sale at €1,000,000.
What that covers: the domain, the site, the extraction pipeline and its output, and the API with the account and billing plumbing behind it. The pipeline rebuilds the whole dataset from public sources on one ordinary machine, with no GPU.
Last updated with 0.9. This describes what the application actually does: every row below is a field you can find in the source. Browsing word pages needs no account, sets no cookie and is not logged against you.
| Data | When | Why | Kept |
|---|---|---|---|
| Email address | you create an account | to identify the account and confirm it | until deletion |
| Password | you create an account | stored only as a scrypt hash, not recoverable, not readable | until deletion |
| API keys | you request or buy one | stored as a SHA-256 hash; the key itself is shown once, at issue | until deletion |
| IP address | any API request | rate limiting, stored as a salted hash, salt rotates daily, raw address never written | rolling window |
| Payment details | you buy a plan | handled entirely by the payment provider: card data never reaches this application | never held here |
Account data and payment records: performance of a contract, plus the record-keeping tax law requires. Hashed IPs: legitimate interest in keeping a free API usable. There is no profiling and no automated decision-making beyond a rate-limit counter.
The hosting provider, the managed database, the payment provider and the email provider that sends confirmation links. No analytics, no advertising network, no third-party trackers: the site loads no script it does not serve itself.
One: a signed session cookie set when you log in, marked
SameSite=Lax and HTTPS-only in production. It is strictly
necessary for logging in, which is why there is no consent banner. Your
theme choice is kept in your browser's local storage and never sent
anywhere.
Access, correction, export, deletion, restriction, objection, and complaint to your national supervisory authority. Ask through the contact above and expect a reply within a month. Being straight about one gap: there is no self-service delete button yet, so deletion requests are carried out by hand.
Passwords are scrypt-hashed, API keys are stored only as hashes, the application runs read-only against the word database, and it holds no card data. That is a small blast radius by design, not an assurance that nothing can go wrong.