Skip to content
Armath
Features Technology Armath Full Stories Support us
ENՀԱՅРУС
Get Armath
Features Technology Armath Full Stories Support us Help Get Armath

Technology and licences

Updated: 24 October 2026

Armath combines project-authored Armenian input intelligence with transparent, offline third-party components. None of these components receives typed text or other user data.

Armath Input Intelligence

Armath’s geometry matcher, candidate decoder, contextual ranking, autocorrection, spelling hints, neural language model and bounded on-device personalisation are written and trained by the Armath project. They work locally. The Armenian word list they search is assembled from the sources credited below.

Armenian word list

Armath’s built-in word list holds the most frequent word forms of the Armenian part of FineWeb-2 by Hugging Face, in order of frequency; it stores the words and their order, not counts or source text. A second, hashed list of further word forms seen in the same corpus only tells the keyboard that a typed word exists. A third list holds further, less frequent word forms of the same corpus that a swipe may also write. FineWeb-2 is made available under the Open Data Commons Attribution License v1.0 and is subject to the Common Crawl Terms of Use. Conversational words keep the position they have in the Armenian frequency list of FrequencyWords by Hermit Dave, generated from the OpenSubtitles corpus (subtitles from OpenSubtitles.org, distributed by OPUS), licensed under CC BY-SA 4.0; the adapted ranking is provided under the same licence.

HySpell Armenian spell-checking dictionary

Which of the rarer word forms the hashed list holds was decided with the reformed-orthography dictionary of HySpell 3.0.1: a form that this dictionary accepts, and that a slip of the finger does not make out of a much more frequent word, is left as typed and is not marked as a spelling error. Armath ships only hashes of those word forms, not the dictionary, its rules or a spell-checking engine. HySpell was created by Haro Mherian and is © 2017 hyspell.com, released under the MIT licence; the complete notice, the author’s memo and the permission notice are included in the app. Use of the HySpell name is attribution only and does not imply sponsorship, affiliation or endorsement.

FUTO Swipe model weights

Powered by FUTO Swipe technology. Armath includes the unmodified FUTO Swipe honorable_sturgeon model weights under the FUTO Model Weights License 1.0. A complete copy of that licence is included in the keyboard extension. Use of the FUTO Swipe name is attribution only and does not imply sponsorship, affiliation or endorsement. FUTO Swipe project

ExecuTorch

Armath uses the BSD-licensed ExecuTorch runtime for on-device model inference. The full BSD copyright notice and disclaimer are included in both the app and the keyboard extension and can be read from Armath’s Settings.

Universal Dependencies Armenian ArmTDP

Aggregate next-word counts are adapted from the train split of Universal Dependencies Armenian ArmTDP. Armath transforms the treebank into aggregate count data and does not ship the source sentences. The adapted data is provided under CC BY-SA 4.0. The ArmTDP and BSUT treebanks are also used, outside the app, to evaluate Armath.

FineWeb-2 Armenian web text

The next-word counts also include aggregate counts of short word sequences from the Armenian part of FineWeb-2 by Hugging Face. Armath’s own next-word language model was trained on the same text. Armath ships only these counts and the trained model, never the source texts. FineWeb-2 is made available under the Open Data Commons Attribution License v1.0 and is subject to the Common Crawl Terms of Use.

GeoNames place names

Selected Armenian names of provinces and major cities are adapted from the GeoNames Armenia extract, licensed under CC BY 4.0.

Noto fonts

The app bundles Noto Serif Armenian, and the letter in the app icon (also the mark of this site) is drawn from Noto Sans Armenian. Both are licensed under the SIL Open Font License 1.1.

This website

This site uses no third-party code, fonts or trackers. Its keyboard drawings, demos and animations are its own; the small in-browser swipe demo uses the 5,000 most frequent words of Armath’s word list.

Armath

The Armenian keyboard for iPhone, powered by Armath Input Intelligence.

Armath

  • Features
  • Technology
  • Armath Full
  • Privacy
  • Questions

Stories

  • Our story
  • The method
  • The feel of the keys
  • Why Armenian
  • Armenian letters
  • Support us

Help

  • Support and setup
  • Privacy policy
  • Technology and licences
  • support@armath.app
Armath
Built in Armenia for Armenian Neural swipe evidence powered by FUTO Swipe technology © 2026 Armath