iito-sama/tsumukey-models

TsumuKey models

Models the TsumuKey iOS keyboard downloads on demand. models.json lists them with sizes and SHA-256 checksums; the app checks the checksum before using a file.

File	What	Size
ranker-r10.rank	Small ranker, round 10 (v10): autocorrection and next-word suggestions (TKR3 format)	4.5 MB
gemma-3-270m-enfrdees64k-Q4_0.gguf	Gemma 3 270M, vocabulary cut to English, French, German, Spanish, Q4_0	102 MB

Small ranker

Scores how well a word fits after the text before it. Words are hashed character n-grams (no vocabulary); a 2-layer transformer reads the last 8 words. Trained on simulated noisy typing of Tatoeba and Wikipedia sentences in the four languages. Mac test set (12,351 words): 95.7% / 87.2% words right typing carefully / fast, top-3 next-word suggestions 29.4% (Gemma 270M: 30.2%); with one language enabled, English 35.4%.

Gemma 270M (modified)

This file is a modified version of Google's Gemma 3 270M: its vocabulary was cut to the four languages above and it was quantized to Q4_0.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms. Use of this model must follow the Gemma Prohibited Use Policy at ai.google.dev/gemma/prohibited_use_policy. See GEMMA_NOTICE.md.