MaskLM

Name what to find.
Get every mention, exactly where it is.

Describe the things you want in plain words. MaskLM returns one object for each, with every span of the text that mentions it, in a single fast pass. Nothing is generated or paraphrased.

Try the playground See how it compares

How it works, step by step

Why MaskLM

The precision of a trained extractor, with the flexibility of a prompt.

More accurate0.77set F1 on documents it never trained on

Ahead of Claude Haiku 5.5 (0.59) and GLiNER (0.41) on the same documents and scorer. See the comparison →

Faster than an LLMMillisecondsper document, not seconds

One forward pass reads the whole document. No token-by-token generation, so a long answer takes no longer than a short one.

Grounded0 hallucinationsevery answer is a span of your document

Nothing is paraphrased or made up: each result points at exact character positions, ready to highlight, mask or cite.

Whole objectsEvery mentiongrouped into one object

"John Lennon", "Lennon" and "he" come back as one person, with a calibrated confidence. No clustering step afterwards.

Private and cheapNo token billa 0.6 B model you can host yourself

Documents never leave your infrastructure, cost doesn't grow with output length, and the same input always gives the same answer.

Any schemaNo retrainingdescribe new types in plain words

Types and fields are text the model reads, so people, deal roles, genes or case citations all work from the same model.

Use cases

Built for the jobs where the exact words matter.

Knowledge graph construction

Turn documents into the graph you designed

Define node types, attributes and relationships. MaskLM finds who plays each role and matches the same entity within and across documents.

Your schema, not a fixed ontologyBuild a graph

Redaction

Hide everything that could identify someone

Names, identifiers and the indirect clues around them, in judgments, contracts or any text: as black boxes, labels or placeholders, with the original a key-press away.

Every mention, even “he” and “the applicant”Redact a document