The smallest ClickBook on-device reading model, and the most thoroughly
tested. Vocabulary pruned to Latin, Cyrillic and Arabic: eleven European
languages plus Arabic.
1.666 GB.
A reader taps a word in a book; the model explains it in the sense that sentence
gives it, writes fresh examples, translates the word, and translates the passage.
Take this one if you serve only these eleven languages. It is 178 MB smaller
than multi at the same quality, and these are the languages with the most
benchmark evidence behind them.
Take multi if you need any Asian or Indic language. The 178 MB it costs buys
seven more languages at no measured quality cost — and note that on Android both
files exceed Play's 1.5 GB asset-pack limit anyway, so the smaller file does not
simplify packaging.
Languages
language
reading
answering
evidence
English, German, Arabic
yes
yes
90-item benchmark, two seeds
French
yes
not yet
183 items, older build: 86.3
Portuguese
yes
not yet
183 items: 85.7
Spanish, Italian
yes
not yet
183 items: 81.3
Russian
yes
not yet
183 items: 81.2
Dutch
yes
not yet
183 items: 79.2
Polish
yes
not yet
183 items: 76.7
Turkish
yes
not yet
183 items: 69.3 — weakest
"Answering" means the model can produce the TRANSLATION and CONTEXT tabs in that
language. Those templates exist for English, German and Arabic only; the other
eight can be read from, with answers in one of those three.
Turkish is the weakest language in the set by a clear margin — agglutinative
morphology fragments hardest under a pruned vocabulary.
Evaluation
90 held-out tapped words in English, German and Arabic, graded 0–100 by an LLM
judge against a rubric containing a reference sense.
build
vocabulary
size
score
same weights at f16
231,955
8.676 GB
80.5
multi (18 languages)
231,955
1.844 GB
79.2
this model
180,850
1.666 GB
79.0 / 77.7 (two seeds)
unpruned vocabulary
262,144
1.950 GB
63.6
This is the only build measured at two seeds, giving 78.4 ± 0.7 — worth
knowing, because single-seed differences of a point or so between these builds are
inside that noise. The apparent 0.2 gap to multi is not meaningful; the 15.6 gap
to the unpruned build is.
Quantisation costs 1.3 points against f16 for a 4.7× smaller file. Pruning further
— to this build's 180,850 tokens — costs nothing measurable.
Without it the model reasons before answering, spends the whole token budget in
reasoning_content, and returns empty content with finish_reason: "length".
That is indistinguishable from a broken model. Raise all caps to 1000 first if you
want reasoning deliberately.
prompts.json here is filtered to this build's eleven languages. The wider file
shipped with multi includes Chinese, Japanese, Korean, Hindi, Tamil, Thai and
Hebrew — sending those prompts to this model would hand it text it has no
tokens for.
Per-tab caps: MEANING 111, EXAMPLE 222, TRANSLATION 111, CONTEXT 444. Measured on
the benchmark, 0 of 360 panels reached them.
Limitations
Not a chat model. Tuned for four narrow tasks driven by the supplied prompts.
Failures concentrate on polysemous common words — bank, charge, Schloss —
15 of 90 items.
No on-device dictionary is bundled. An external sense resource measured +12.7
points on this benchmark but is not included.
Modifications, as Apache 2.0 requires derivative works to state:
Vocabulary pruned from 262,144 to 180,850 tokens, retaining Latin, Cyrillic
and Arabic and removing CJK, Hangul, Devanagari, Thai, Hebrew, Greek and other
scripts the product does not serve.
Quantised to IQ4_XS with Q2_K token embeddings under an importance matrix.
No weights were fine-tuned, distilled or retrained.
Gemma is a trademark of Google LLC. This is an independent derivative, not
endorsed by or affiliated with Google.
