Clickbook/ClickBook-Gemma-4-E2B-multi-IQ4_XS

ClickBook Gemma 4 E2B — multilingual, vocabulary-pruned, IQ4_XS

An on-device reading assistant model. 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.

1.844 GB, built to fit a 1.85 GB application budget on iOS and Android.

What was changed from the base model

1. Vocabulary pruned from 262,144 to 231,955 tokens. Scripts not served by the product were removed: Bengali, Gurmukhi, Gujarati, Oriya, Telugu, Kannada, Malayalam, Sinhala, Lao, Tibetan, Myanmar, Georgian, Ethiopic, Cherokee and others. Latin, Cyrillic, Arabic, CJK, Hangul, Devanagari, Tamil, Thai, Greek and Hebrew are retained.

   This matters more than it would in most architectures: Gemma 4 carries per-layer embeddings, so a vocabulary token costs 10,496 parameters — about 3.49 KB in this quantisation. Vocabulary is roughly a third of the file.

2. Quantised to IQ4_XS with Q2_K token embeddings, guided by an importance matrix computed over 1,538 calibration chunks.

   Embedding tensors are deliberately left out of imatrix guidance because they cannot benefit from it — embedding lookups are gather operations, not matrix multiplications, so llama-imatrix never observes them.

Files

file	purpose
ClickBook-Gemma-4-E2B-multi-IQ4_XS.gguf	the model
prompts.json	every prompt, per language, with placeholders and token caps
MOBILE-INTEGRATION.md	integration guide for client developers

Usage

llama-server -m ClickBook-Gemma-4-E2B-multi-IQ4_XS.gguf -c 2048 -ngl 99 --jinja

{
  "messages": [{ "role": "user", "content": "<prompt from prompts.json>" }],
  "temperature": 0.1, "top_k": 40, "top_p": 0.9, "repeat_penalty": 1.05,
  "max_tokens": 111,
  "chat_template_kwargs": { "enable_thinking": false }
}

enable_thinking: false is required

Without it the model reasons before answering, spends the entire token budget in reasoning_content, and returns empty content with finish_reason: "length". That is indistinguishable from a broken model or an unsupported language. If you want reasoning, raise every cap to 1000 first.

Language support

Support is uneven and the model card should be believed over the language tag list above, which cannot express degrees.

languages	reading	answering	evidence
English, German, Arabic	yes	yes	90-item benchmark, two seeds
French, Portuguese, Spanish, Italian, Russian, Dutch, Polish, Turkish	yes	not yet	183 items each on an earlier build, 69–86
Chinese, Japanese, Korean, Hindi, Thai, Hebrew	yes	yes	smoke-tested, 4/4 tabs pass
Tamil	yes	no	passage translation into Tamil is corrupted
Greek	no	no	output mixes scripts and invents words

Greek is broken. Do not enable it. Only 1,511 Greek tokens survived the prune, which is not enough for coherent generation; output splices Cyrillic letters and German words into Greek text.

Tamil should be read but not written to. Translating a passage into Tamil produced Han, Kana and Hangul characters spliced into Tamil script.

The six smoke-tested languages passed every automated check and read two deliberately hard idiomatic taps each correctly — 手が空いた, 발이 넓다, हाथ धोना, ใจดี, יד חמה. That is a handful of sentences, not a benchmark. Treat them as beta.

Evaluation

Measured on a 90-item held-out set of tapped words in English, German and Arabic, graded 0–100 by an LLM judge against a rubric containing a reference sense. Same prompts, sampler and seed throughout; the only variable is the model file.

build	vocabulary	size	score	failures
same weights at f16	231,955	8.676 GB	80.5	12
this model	231,955	1.844 GB	79.2	15
narrower prune, 11 languages	180,850	1.666 GB	79.0	15
unpruned vocabulary, same quantisation	262,144	1.950 GB	63.6	23

Three things this shows.

Quantisation is nearly free. IQ4_XS with Q2_K embeddings costs 1.3 points against the same weights at f16, for a 4.7× smaller file — 8.676 GB to 1.844 GB.

Restoring seven scripts is free. Going from 180,850 to 231,955 tokens — adding CJK, Hangul, Devanagari, Tamil, Thai, Greek and Hebrew — moved the score by 0.2 and the failure count not at all.

Pruning is not only a size optimisation, it is a quality one. The unpruned 262,144-token vocabulary, built with an identical recipe and imatrix, scores 15.6 points lower and fails 23 items instead of 15. The regression is worst on the easiest third of the set (91.5 → 70.8). A plausible mechanism is that at Q2_K precision the embedding table and output distribution spend capacity on tens of thousands of tokens the model never needs, at the expense of the tokens it does. The measurement is single-seed and reported as measured; the explanation is a hypothesis.

Packaging note for mobile

The file compresses by only 3% (ratio 0.970) — quantised weights are close to incompressible, so download size is effectively file size. Google Play's asset pack limit is 1.5 GB per pack, which this model exceeds, so Android delivery needs two asset packs concatenated on device. Every smaller variant above exceeds it too, so splitting is unavoidable regardless of which build is chosen.

Limitations

• Not a chat model. It is tuned for four narrow tasks driven by the supplied prompts. General conversation is out of scope.
• The hardest word senses remain hard: polysemous common words (bank, charge, Schloss) are where the 15 failures concentrate.
• No on-device dictionary is bundled. An external sense resource was measured at +12.7 points on the same benchmark but is not included here.
• Quality in the newly-restored scripts is established by smoke tests only.

License and provenance

Apache License 2.0, matching the licence declared by the base model, google/gemma-4-E2B-it.

Google additionally publishes a Gemma 4 license page, linked from the upstream model card; consult it directly for anything the Apache 2.0 grant does not answer.

Modifications from the base model

Apache 2.0 requires derivative works to state their changes. Two were made, both described in full above:

1. Vocabulary pruned from 262,144 to 231,955 tokens, removing scripts the product does not serve.
2. Quantised to IQ4_XS with Q2_K token embeddings, guided by an importance matrix.

No weights were fine-tuned, distilled or otherwise retrained. Model architecture and all non-vocabulary tensors are the base model's, requantised.

Gemma is a trademark of Google LLC. This is an independent derivative, not endorsed by or affiliated with Google.