Gemma4GR v3: Gemma 4 E4B fine-tuned for cleaner Greek (stray foreign characters in ~1% of answers vs ~40% for stock). v2 tags were the base model (export bug), now fixed. Not a source of facts.

tools thinking

ollama run efso/gemma-4-e4b-it-gr-v2

curl http://localhost:11434/api/chat \
  -d '{
    "model": "efso/gemma-4-e4b-it-gr-v2",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='efso/gemma-4-e4b-it-gr-v2',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'efso/gemma-4-e4b-it-gr-v2',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Applications

Claude Code ollama launch claude --model efso/gemma-4-e4b-it-gr-v2

OpenCode ollama launch opencode --model efso/gemma-4-e4b-it-gr-v2

Hermes Agent ollama launch hermes --model efso/gemma-4-e4b-it-gr-v2

OpenClaw ollama launch openclaw --model efso/gemma-4-e4b-it-gr-v2

Models

View all →

Name

4 models

Size / Usage

Context

Input

gemma-4-e4b-it-gr-v2:latest

5.3GB · 128K context window · Text · 1 month ago

gemma-4-e4b-it-gr-v2:latest

5.3GB

128K

Text

gemma-4-e4b-it-gr-v2:q4_k_m

5.3GB · 128K context window · Text · 1 month ago

gemma-4-e4b-it-gr-v2:q4_k_m

5.3GB

128K

Text

gemma-4-e4b-it-gr-v2:q6_k

6.2GB · 128K context window · Text · 1 month ago

gemma-4-e4b-it-gr-v2:q6_k

6.2GB

128K

Text

gemma-4-e4b-it-gr-v2:q8_0

8.0GB · 128K context window · Text · 1 month ago

gemma-4-e4b-it-gr-v2:q8_0

8.0GB

128K

Text

Readme

Gemma4GR E4B v3

Gemma 4 E4B with a Greek Q&A fine-tune merged in. Tags: latest (= q4_k_m), q4_k_m, q6_k, q8_0. Text-only; for audio or image input use the GGUFs and mmproj from Hugging Face.

Correction: the earlier v2 tags on this page contained the unmodified base model google/gemma-4-E4B-it, because of an export bug. v3 is the first build with the fine-tune actually merged, checked tensor by tensor against the base.

Measured against stock Gemma 4 E4B under identical settings (150 held-out Greek questions): stray foreign-script fragments in about 1% of answers vs about 40%; it stops properly instead of rambling; answers are closer to references (chrF 0.39 vs 0.30).

Limitations: it does not know more than the base model and states wrong facts confidently. Treat it as a Greek language model, not a source of facts.

Details, method and checksums: https://huggingface.co/Efso/gemma-4-E4B-it-GR-v2