Instructions to use Mediform/gemma4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Mediform/gemma4 with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="Mediform/gemma4",
	filename="gemma-4-E4B-it-ckpt500-Q4_K_M.gguf",
)

llm.create_chat_completion(
	messages = [
		{
			"role": "user",
			"content": "What is the capital of France?"
		}
	]
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use Mediform/gemma4 with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Mediform/gemma4:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Mediform/gemma4:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf Mediform/gemma4:Q4_K_M
# Run inference directly in the terminal:
llama-cli -hf Mediform/gemma4:Q4_K_M

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf Mediform/gemma4:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf Mediform/gemma4:Q4_K_M

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf Mediform/gemma4:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf Mediform/gemma4:Q4_K_M

Use Docker

docker model run hf.co/Mediform/gemma4:Q4_K_M

LM Studio
Jan

vLLM

How to use Mediform/gemma4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Mediform/gemma4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Mediform/gemma4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Mediform/gemma4:Q4_K_M

Ollama
How to use Mediform/gemma4 with Ollama:
```
ollama run hf.co/Mediform/gemma4:Q4_K_M
```

Unsloth Studio

How to use Mediform/gemma4 with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Mediform/gemma4 to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for Mediform/gemma4 to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for Mediform/gemma4 to start chatting

Docker Model Runner
How to use Mediform/gemma4 with Docker Model Runner:
```
docker model run hf.co/Mediform/gemma4:Q4_K_M
```

Lemonade

How to use Mediform/gemma4 with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull Mediform/gemma4:Q4_K_M

Run and chat with the model

lemonade run user.gemma4-Q4_K_M

List all available models

lemonade list

gemma-4-E4B-it — text-finetuned (QLoRA) — GGUF

GGUF export of a QLoRA fine-tune of gemma-4-E4B-it, quantized to Q4_K_M for llama.cpp. The fine-tune targets text generation (German medical domain); the multimodal audio (and vision) capabilities are inherited unchanged from the base model.

Files

File	Size	Description
`gemma-4-E4B-it-ckpt500-Q4_K_M.gguf`	5.0 GB	Language model. Base gemma-4-E4B-it with the text-generation QLoRA adapter (checkpoint-500) merged in, quantized to Q4_K_M.
`mmproj-gemma-4-E4B-it-BF16.gguf`	946 MB	Multimodal projector (BF16) — audio + vision encoders. Required for audio/image input. Not fine-tuned.

⚠️ Chat template / system prompt (important)

The base gemma-4-E4B-it is a reasoning model: its original chat template auto-enables a <channel>thought reasoning trace whenever the first message has role system or developer. For a JSON-only extraction task this looks like "gibberish" (the model emits a thinking trace instead of the JSON). This GGUF ships a corrected chat template that folds any system/developer message into the first user turn and never enables thinking — so the model answers directly. Just deploy normally with --jinja:

# Text + audio
llama-server -m gemma-4-E4B-it-ckpt500-Q4_K_M.gguf \
             --mmproj mmproj-gemma-4-E4B-it-BF16.gguf --jinja

# Text only
llama-cli -m gemma-4-E4B-it-ckpt500-Q4_K_M.gguf --jinja -st -sysf system.txt -f user.txt

system + user chat-completions requests now return the trained JSON directly. (If you ever use the base model's template instead, send the instructions in the user message — do not use a system role — to avoid the reasoning trace.)

What is and isn't fine-tuned

✅ Text generation — QLoRA adapter (rank 8, checkpoint-500) merged into the language model.
➖ Audio / vision — base mmproj encoders, unchanged.

Provenance

Base model: unsloth/gemma-4-E4B-it (16-bit).
Adapter: LoRA r=8, α=8, checkpoint-500.
Export: adapter → GGUF (convert_lora_to_gguf.py) → merged into base text GGUF (llama-export-lora) → quantized Q4_K_M (llama-quantize) → chat template corrected (gguf_new_metadata.py).

Downloads last month: 56

GGUF

Model size

8B params

Architecture

gemma4

Hardware compatibility

4-bit

Model tree for Mediform/gemma4

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Finetuned

unsloth/gemma-4-E4B-it

Quantized

(13)

this model