Gemma 4 E2B Español compact for Fluent Books (light-2, IQ4_XS GGUF)

The light model of Fluent Books that translates the sentences of a Spanish novel into Russian, on the device, for devices with less memory than the full one needs. 967,251,872 bytes (IQ4_XS with the per-layer projection in Q8_0), SHA-256 52e7c42dd1b11b1e629e5bdd284780c6b4845afe7f99dadc640c3690d5ca3b3f.

It is not a chat model: it answers the app's "translate this sentence from a novel" and nothing else.

How it was made

1. ES2: google/gemma-4-E2B-it with its vocabulary cut to 48,000 tokens, LoRA fine-tuned on 38,336 rows (sentences of public-domain Spanish novels translated by larger Gemma and Claude agents, and checked rows for traps, set expressions, collocations and names).
2. The feed-forward layers of all 35 layers narrowed to 6,144 neurons (the ones kept have the largest activation), all layers kept (taking layers out cost the expressions and names).
3. Distilled from ES2 on the same 38,336 rows (2 epochs, cross-entropy plus KL to ES2's logits).
4. Quantized to IQ4_XS with an importance matrix from the training prompts; the per-layer projection kept in Q8_0.

What it does, measured (our own sets, 6–7 Nov 2026; two judges, a translation counts only if both mark it right)

Memory (llama.cpp, CPU, a 2,048-token context, one sentence): about 1.4 GB. On a device with Metal not measured.

What it does not do well

• Idioms, proverbs and names that are also words: halving the feed-forward width costs them. A student distilled from the better teacher (ES3b) on a third of the data was worse still (390 against 421 of 684), so this one stays.
• Slang, regional words, long sentences (a check of 31 slang sentences: about a quarter right for the full model; not measured for this one).
• Only the sentence is translated; no choice of word meanings.
• It can fall into Latin letters inside a Russian word; the app forbids tokens of plain Latin letters in the answer (our measurements used the same ban).
• Our sets are small and ours; the judges were language-model agents.

Training data

• Sentences from public-domain Spanish novels (Project Gutenberg), translated by Claude and Gemma agents.
• Wiktionary and Wikipedia (CC BY-SA) as a source of phrases, forms and word frequencies.
• Phrase lists and labels from licensed dictionaries and idiom books owned by the authors. The books' text is not distributed here.

Use

llama.cpp (a build that knows Gemma 4). Gemma 4 turns, no thinking:

<bos><|turn>system
You are a literary translator from Spanish into Russian.<turn|>
<|turn>user
Translate this sentence from a novel into natural Russian. Keep its meaning exactly; output only the translation.

<the Spanish sentence><turn|>
<|turn>model

Base model: Google's Gemma 4 E2B (Apache-2.0); thanks to the Gemma team.