OpenSwipe 0.6.0-1

Installed before 30 October 2026 (older than version 0.5.1-1)? Install it once more (how). Older versions can't find updates after the move to the new name. After that it updates itself.

Arch/CachyOS x86_64, Fedora 44 x86_64/aarch64 (ARM64 Linux Plasma Mobile), Kubuntu 26.10, SteamOS/Bazzite and older KDE (Flatpak) x86_64; Debian testing and openSUSE Tumbleweed when their build passed. Complete corresponding source and SHA256SUMS included. CI checks package/build/loader contracts; physical compositor/input acceptance is separate and must identify package hashes. ARM session acceptance is required; emulated ARM input does not certify physical-phone performance. Android, iOS, musl and Gaming Mode are not certified.

Install and enable · Release scope

Swipe Keyboard 0.5.1-1

Arch/CachyOS x86_64, Fedora 44 x86_64/aarch64 (ARM64 Linux Plasma Mobile) and SteamOS Desktop Flatpak x86_64. Complete corresponding source and SHA256SUMS included. CI checks package/build/loader contracts; physical compositor/input acceptance is separate and must identify package hashes. ARM session acceptance is required; emulated ARM input does not certify physical-phone performance. Android, iOS, musl and Gaming Mode are not certified.

Install and enable · Release scope

Swipe Keyboard 0.5.0-1

Swipe Keyboard 0.5.0-1: swipe typing, word suggestions and spelling help for KDE Plasma's on-screen keyboard (English US/UK, Spanish, French, German). Runs on your device; typing recording and learning are off by default. Test release for KDE Plasma on Wayland, plus a preview of Steam Deck Gaming Mode.

New in 0.5.0

• No FUTO code or models. Swipe recognition, suggestions and learning are the project's own open work. Models trained and released by Destin Moss via youcoded.ai, 2026. Licences: MIT where possible; KDE's GPL choice only for the two keyboard add-on layers.
• Plasma 6.9 support. One package carries the keyboard for Plasma 6.7/6.8 and for 6.9's rebuilt keyboard, and starts the right one. The 6.9 version is tested against KDE's development code, not yet on a real Plasma 6.9 device.
• Better words: "we're" or "were" picked from the sentence, capital letters for "I", names, places, days and months (English, Spanish, French), texting words (wyd, lmao, jaja, mdr ...), and cleaner other guesses after a swipe: stretched spellings, rare words and rare names fell from 44 % of them to 5 % (English).
• Tap a word you swiped to swap it for one of its alternatives, now also in ordinary apps such as browsers, for words the keyboard typed in that field.
• Learning from your typing (off by default; swipe-keyboard learning on).

Honest limits

• Swipe typing is a few points less accurate than the FUTO engine used up to 0.4.1 in English, Spanish and German. English, 1,946 real swipes inside sentences: first guess 91.5 % (FUTO 93.3 %), right word in the top 3 94.6 % (96.6 %); the top-3 drop is rare words and names, common words gained slightly. Without a sentence: Spanish 75.9 % (77.7 %), German 64.7 % (69.7 %); French is better, 68.2 % (64.6 %).
• Bigger Spanish, French and German word lists, the measured way to lift those numbers, are still being tested and will come in a later update.
• Next-word suggestions are ahead of Gboard on every recorded measure in all four languages (Gboard 14.2, Android 35 emulator, learning off, tested 2026-10-22). Nothing is claimed about newer Gboard versions or Gboard with learning on. Swipe typing was not compared with Gboard.

Updating: 0.4.1 installs this update by itself within a day (off switch: sudo swipe-keyboard updates off, or bash install.sh --no-auto-update on the Deck). 0.4.0 and older: install once by hand.

Pick your file: Arch/CachyOS x86_64 (.pkg.tar.zst), Fedora 44 x86_64 or ARM64 (.rpm), Steam Deck Desktop Mode (.flatpak), Steam Deck including Gaming Mode (-steamos-x86_64.tar.gz). Step-by-step: README. Check downloads with SHA256SUMS.

Good to know: complete corresponding source is included. Package builds are checked by CI; physical-device acceptance is separate, and ARM results from emulation do not certify phone performance. Gaming Mode is a preview tested on a virtual Deck only. Not certified: Android, iOS, musl distributions.

Release scope and safeguards · Models and how they were measured

Swipe Keyboard 0.4.1-1

Swipe Keyboard 0.4.1-1: swipe typing, word suggestions and spelling help for KDE Plasma's on-screen keyboard (English US/UK, Spanish, French, German). Runs on your device; typing recording is off by default. Test release for KDE Plasma on Wayland, plus a preview of Steam Deck Gaming Mode.

New in 0.4.1

• Updates install themselves once a day on every platform, accepting only updates signed with the project key. 0.4.0 and older must install 0.4.1 once by hand.
• Gaming Mode preview now works with Steam's stable client (tested on a virtual Steam Deck only, not yet a real one).

Pick your file: Arch/CachyOS x86_64 (.pkg.tar.zst), Fedora 44 x86_64 or ARM64 (.rpm), Steam Deck Desktop Mode (.flatpak), Steam Deck including Gaming Mode (-steamos-x86_64.tar.gz). Step-by-step: README. Check downloads with SHA256SUMS.

Good to know: swipe recognition in this version is FUTO's engine, under FUTO's licence and notices (version 0.5.0 replaces it with the project's own). Complete corresponding source is included. Package builds are checked by CI; physical-device acceptance is separate, and ARM results from emulation do not certify phone performance. Not certified: Android, iOS, musl distributions.

Release scope and safeguards

Language files 2026-10-23, apostrophe-aware (en, es, fr, de)

Language files, not a keyboard update. The files the Swipe Keyboard 0.5.0 uses for English, Spanish, French and German: swipe dictionaries, word lists and word-prediction (n-gram) files. Older keyboard versions keep working with their previous release; they will not load these files (each version pins its own checksums).

• What changed from language-files-2026-10-22-open-words: en/es/fr/de.ngram and en/es/fr/de.vocab.tsv now keep apostrophes, so "were" and "we're", "its" and "it's", "well" and "we'll", "lune" and "l'une" are separate words and the keyboard can pick the one that fits the sentence. Everything else (swipe dictionaries, word counts, credits, notice and licence texts) is byte-for-byte the same.
• Measured on 2,000 held-out sentences per language: English apostrophe-pair words swiped right first time 78.0 % to 87.7 %, in the top 3 81.9 % to 96.8 %; French pairs 86.3 % to 91.1 %; all English words 94.0 % to 94.4 %. Next-word and completion scores, including the comparison with Gboard 14.2 (Android 35 emulator, learning off, tested 2026-10-22), moved by at most 0.2 points.
• Sources and licences: FineWeb / FineWeb-2 web text (ODC-By 1.0), Tatoeba sentences (CC BY 2.0 France), Android's AOSP word list (Apache-2.0) for the English dictionary. Attribution only, no share-alike. Keep NOTICE.md, *-CONTRIBUTORS.txt and the licence texts with the files.
• Checksums: SHA256SUMS.

Swipe Keyboard 0.4.0-1

Arch/CachyOS x86_64, Fedora 44 x86_64/aarch64 (ARM64 Linux Plasma Mobile) and SteamOS Desktop Flatpak x86_64. Complete corresponding source and SHA256SUMS included. CI checks package/build/loader contracts; physical compositor/input acceptance is separate and must identify package hashes. ARM session acceptance is required; emulated ARM input does not certify physical-phone performance. Android, iOS, musl and Gaming Mode are not certified.

Install and enable · Release scope

Trained models 2026-10-23 (next-word en/es/fr/de, swipe)

Trained models, not a keyboard update. Models trained and released by Destin Moss via youcoded.ai, 2026.

Five small models that run on the user's own device inside the Swipe Keyboard (swipe typing and word suggestions for KDE Plasma's on-screen keyboard; English, Spanish, French, German). Nothing is sent anywhere.

File	What it does	Size	Licence
en.nwp.bin, es.nwp.bin, fr.nwp.bin, de.nwp.bin	Neural next-word models that read the sentence so far. Used together with the n-gram files from language-files-2026-10-23-apostrophes.	about 4.7 MB each	ODC-By-1.0 AND CC-BY-2.0-FR (attribution only; trained on FineWeb/FineWeb-2 and Tatoeba)
swipe-model.bin	Turns a finger path into ranked words; one file for all four languages.	2.1 MB	MIT (trained on FUTO's MIT swipe.futo.org data plus our synthetic swipes; no FUTO code or model weights)

How good: on our recorded test queries the next-word models plus their n-gram files were ahead of Gboard on every measure, by 3.9 to 19.6 points. The test: Gboard 14.2 (Android 35 x86_64 emulator image), learning off, tested 2026-10-22, 100 sentences per set. Newer Gboard versions and Gboard with learning on were not measured. On English swipes the keyboard with this swipe model is about 2 points behind the FUTO engine it replaces (91.5 % against 93.3 % first guess on the same 1,946 real swipes, earlier words of the sentence known). Model cards with every table: https://huggingface.co/youcoded. How it was measured: https://huggingface.co/youcoded/swipe-typing/blob/main/METHODS.md

To use them: the keyboard downloads and checks every file against a pinned checksum (SHA256SUMS); a missing or altered model only turns that model off. Developers: start with swipekb::LanguageEngine (a developer guide, engine/README.md, arrives with version 0.5.0).

Credits and sources: NOTICE.md. Keep ODC-By-1.0.txt and TATOEBA-CC-BY-2.0-FR.txt with the next-word models and FUTO-SWIPE-DATA-LICENSE.txt with the swipe model.

Language files 2026-10-22, open word lists (en, es, fr, de)

Language files, not a keyboard update. Word lists, swipe dictionaries and word-prediction (n-gram) files for English, Spanish, French and German, counted from openly licensed text. Superseded for the keyboard by language-files-2026-10-23-apostrophes, which rebuilds the n-gram and word-list files; the other files here are unchanged in that release. The next-word models in models-2026-10-23 and the published accuracy tables were measured against these files.

• Sources: FineWeb / FineWeb-2 web text (ODC-By 1.0), Tatoeba sentences (CC BY 2.0 France) and Android's AOSP word list (Apache-2.0, base of the English swipe dictionary). No FUTO data or code, no wordfreq data. Replaces language-files-2026-10-22, whose word lists came from wordfreq (CC BY-SA); these carry attribution-only terms.
• Included: the swipe dictionaries (*.combined), the word counts they are built from (so they can be rebuilt), n-gram files, word lists, credits and licence texts.
• Measured against the previous files: no measure worse by more than 0.1 point; French and German suggestions improve slightly; German nouns keep their capital letter.
• Credits to keep: NOTICE.md, *-CONTRIBUTORS.txt (Tatoeba authors), licence texts. Checksums: SHA256SUMS.

Swipe Keyboard 0.3.1-1

Arch/CachyOS x86_64, Fedora 44 x86_64/aarch64 (ARM64 Linux Plasma Mobile) and SteamOS Desktop Flatpak x86_64. Complete corresponding source and SHA256SUMS included. CI checks package/build/loader contracts; physical compositor/input acceptance is separate and must identify package hashes. ARM session acceptance is required; emulated ARM input does not certify physical-phone performance. Android, iOS, musl and Gaming Mode are not certified.

Install and enable · Release scope

Swipe Keyboard 0.3.0-1

Arch/CachyOS x86_64, Fedora 44 x86_64/aarch64 (ARM64 Linux Plasma Mobile) and SteamOS Desktop Flatpak x86_64. Complete corresponding source and SHA256SUMS included. CI checks package/build/loader contracts; physical compositor/input acceptance is separate and must identify package hashes. ARM session acceptance is required; emulated ARM input does not certify physical-phone performance. Android, iOS, musl and Gaming Mode are not certified.

Install and enable · Release scope