Instructions to use JMingo/gemma-4-E4B-it-qat-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use JMingo/gemma-4-E4B-it-qat-GGUF with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16
# Run inference directly in the terminal:
llama cli -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16

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 JMingo/gemma-4-E4B-it-qat-GGUF:BF16
# Run inference directly in the terminal:
./llama-cli -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16

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 JMingo/gemma-4-E4B-it-qat-GGUF:BF16
# Run inference directly in the terminal:
./build/bin/llama-cli -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16

Use Docker

docker model run hf.co/JMingo/gemma-4-E4B-it-qat-GGUF:BF16

LM Studio
Jan
Ollama
How to use JMingo/gemma-4-E4B-it-qat-GGUF with Ollama:
```
ollama run hf.co/JMingo/gemma-4-E4B-it-qat-GGUF:BF16
```
Unsloth Desktop

How to use JMingo/gemma-4-E4B-it-qat-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "JMingo/gemma-4-E4B-it-qat-GGUF:BF16"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use JMingo/gemma-4-E4B-it-qat-GGUF with Docker Model Runner:
```
docker model run hf.co/JMingo/gemma-4-E4B-it-qat-GGUF:BF16
```

Lemonade

How to use JMingo/gemma-4-E4B-it-qat-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull JMingo/gemma-4-E4B-it-qat-GGUF:BF16

Run and chat with the model

lemonade run user.gemma-4-E4B-it-qat-GGUF-BF16

List all available models

lemonade list

Hermes Agent

How to use JMingo/gemma-4-E4B-it-qat-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default JMingo/gemma-4-E4B-it-qat-GGUF:BF16

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use JMingo/gemma-4-E4B-it-qat-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf JMingo/gemma-4-E4B-it-qat-GGUF:BF16

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "JMingo/gemma-4-E4B-it-qat-GGUF:BF16" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Near-lossless GGUF quants of:

Uses the chat template from google/gemma-4-E4B-it (2026-05-18), replacing the outdated chat template that was originally bundled with the base QAT model.

Original QAT Specifications

The base QAT model was trained with the following specifications:

Text & Draft Models: 4-bit
Multimodal Projector (mmproj):
- Audio Encoder: Mostly 2-bit (with some tensors in 4-bit or BF16)
- Vision Encoder: 8-bit
- Other parts: BF16

Quantization Error Evaluation

Dequantization: Quantized weights were dequantized to F32 for evaluation.
Error Calculation: Quantization error metrics (compared to the BF16 baseline) were computed and accumulated in F64 precision to prevent numerical underflow and precision loss.

Metrics:

MAE (Mean Absolute Error)
RMSE (Root Mean Squared Error)
Max Error

Interpretation of Metrics: These error metrics do not directly reflect actual model performance. This is because imatrix optimization intentionally increases measured error by downweighting less important parameters to improve overall performance. However, these metrics remain useful for estimating how close the quantized model is to the original, lossless quality.

gemma-4-E4B-it-qat

Model	MAE	RMSE	Max Error
`JMingo/gemma-4-E4B-it-qat-BF16.gguf` (Baseline)	0.00000000	0.00000000	0.00000000
`JMingo/gemma-4-E4B-it-qat-Q4_0.gguf`	0.00002867	0.00005571	0.00256348
`idkwhattoputherenow/gemma-4-E4B-it-qat-q4_0-unquantized-q4_0-maxerr.gguf`	0.00003258	0.00006382	0.00299072
`unsloth/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf`	0.00004080	0.00007184	0.00341797
`google/gemma-4-E4B_q4_0-it.gguf`	0.00046331	0.00084654	0.07275391
`lmstudio-community/gemma-4-E4B-it-QAT-Q4_0.gguf`	0.00046331	0.00084654	0.07275391
`mradermacher/gemma-4-E4B-it-qat-q4_0-unquantized.i1-Q4_0.gguf`	0.00024689	0.00053073	0.03417969
`mradermacher/gemma-4-E4B-it-qat-q4_0-unquantized.Q4_K_S.gguf`	0.00048946	0.00078925	0.04432869
`mradermacher/gemma-4-E4B-it-qat-q4_0-unquantized.Q4_K_M.gguf`	0.00047553	0.00078097	0.04432869
`mradermacher/gemma-4-E4B-it-qat-q4_0-unquantized.i1-Q4_1.gguf`	0.00042801	0.00076768	0.21069336
`mradermacher/gemma-4-E4B-it-qat-q4_0-unquantized.i1-Q4_K_S.gguf`	0.00048423	0.00078383	0.21218872
`mradermacher/gemma-4-E4B-it-qat-q4_0-unquantized.i1-Q4_K_M.gguf`	0.00047063	0.00077560	0.21218872

mtp-gemma-4-E4B-it-qat

Model	MAE	RMSE	Max Error
`JMingo/mtp-gemma-4-E4B-it-qat-BF16.gguf` (Baseline)	0.00000000	0.00000000	0.00000000
`JMingo/mtp-gemma-4-E4B-it-qat-Q4_0.gguf`	0.00005023	0.00008237	0.00329590
`unsloth/mtp-gemma-4-E4B-it.gguf`	0.00007141	0.00010650	0.00854492

mmproj-gemma-4-E4B-it-qat

Model	MAE	RMSE	Max Error
`JMingo/mmproj-gemma-4-E4B-it-qat-BF16.gguf` (Baseline)	0.00000000	0.00000000	0.00000000
`JMingo/mmproj-gemma-4-E4B-it-qat-Q2_K_XL.gguf`	0.00001205	0.00003343	0.00271225

Downloads last month: 775

GGUF

Model size

7B params

Architecture

gemma4

Hardware compatibility

4-bit

16-bit

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for JMingo/gemma-4-E4B-it-qat-GGUF

Base model

google/gemma-4-E4B

Finetuned

google/gemma-4-E4B-it

Finetuned

google/gemma-4-E4B-it-qat-q4_0-unquantized

Quantized

(36)

this model