gemma4-e4b-nothink-gguf

gemma4-e4b-nothink (GGUF, Q4_K_M)

A repackaging of the Gemma 3n E4B Instruct model (8.0B effective params, multimodal: text + vision + audio) as a Q4_K_M GGUF for use with Ollama / llama.cpp, with a "nothink" recipe: chat templating is disabled and the model receives the raw prompt (TEMPLATE {{ .Prompt }} + PARSER passthrough), so it never emits hidden reasoning / <think> blocks. Good for deterministic completion-style use, programmatic prompting, and pipelines that want full control over message framing.

• Architecture: gemma4 (a.k.a. Gemma 3n)
• Parameters: ~8.0B
• Context length: 131,072
• Embedding length: 2,560
• Quantization: Q4_K_M
• File size: ~8.9 GB
• Capabilities: text completion, vision, audio
• Origin: Exported from local Ollama tag gemma4:e4b-nothink

The underlying weights are Google's Gemma model. Use is governed by the Gemma Terms of Use and the Gemma Prohibited Use Policy. By downloading or using these files you agree to those terms.

---

Files

---

Quick start with Ollama

1. Install Ollama

See https://ollama.com/download. Make sure ollama is on your PATH and the daemon is running (ollama serve or the desktop app).

2. Pull the files from this repo

# Install once
pip install -U "huggingface_hub[cli]"

# Download into ./gemma4-e4b-nothink/
hf download tawatchai/gemma4-e4b-nothink-gguf \
  --local-dir ./gemma4-e4b-nothink

3. Register the model with Ollama

cd ./gemma4-e4b-nothink
ollama create gemma4:e4b-nothink -f Modelfile

ollama create reads Modelfile, finds the relative FROM ./gemma4-e4b-nothink.Q4_K_M.gguf, and imports it into Ollama's local store. After this completes you can delete the working directory if you want — the blob is copied into ~/.ollama/.

Verify:

ollama list | grep gemma4
ollama show gemma4:e4b-nothink

4. Run it

Interactive:

ollama run gemma4:e4b-nothink

One-shot:

ollama run gemma4:e4b-nothink "Translate to German: Where is the train station?"

OpenAI-compatible API (default http://localhost:11434/v1):

curl http://localhost:11434/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{
    "model": "gemma4:e4b-nothink",
    "messages": [{"role": "user", "content": "Hello!"}],
    "temperature": 0.1
  }'

---

Quick start with llama.cpp

# Download just the GGUF
hf download tawatchai/gemma4-e4b-nothink-gguf \
  gemma4-e4b-nothink.Q4_K_M.gguf --local-dir .

# Run
./llama-cli -m ./gemma4-e4b-nothink.Q4_K_M.gguf \
  -p "Hello!" --temp 0.1 --top-k 64 --top-p 0.95

Because this is a "nothink" / passthrough variant, you are responsible for any chat formatting (turn markers, system messages, etc.). If you want the standard Gemma chat template instead, use google/gemma-3n-E4B-it directly, or replace TEMPLATE {{ .Prompt }} in the Modelfile with the Gemma chat template.

---

Sampling defaults (from Modelfile)

temperature  0.1
top_k        64
top_p        0.95

Override per-request via the Ollama / OpenAI API as usual.

---

Faster downloads (optional)

pip install hf_transfer
export HF_HUB_ENABLE_HF_TRANSFER=1
hf download tawatchai/gemma4-e4b-nothink-gguf --local-dir ./gemma4-e4b-nothink

---

License

Gemma is provided under the Gemma Terms of Use. See LICENSE in this repo and https://ai.google.dev/gemma/terms.