Instructions to use minas2025/Gemma-4-E2B-Coder-Uncensored with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use minas2025/Gemma-4-E2B-Coder-Uncensored 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 minas2025/Gemma-4-E2B-Coder-Uncensored
# Run inference directly in the terminal:
llama cli -hf minas2025/Gemma-4-E2B-Coder-Uncensored

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf minas2025/Gemma-4-E2B-Coder-Uncensored
# Run inference directly in the terminal:
llama cli -hf minas2025/Gemma-4-E2B-Coder-Uncensored

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf minas2025/Gemma-4-E2B-Coder-Uncensored
# Run inference directly in the terminal:
./build/bin/llama cli -hf minas2025/Gemma-4-E2B-Coder-Uncensored

Use Docker

docker model run hf.co/minas2025/Gemma-4-E2B-Coder-Uncensored

Llama new
LM Studio
Jan

vLLM

How to use minas2025/Gemma-4-E2B-Coder-Uncensored with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "minas2025/Gemma-4-E2B-Coder-Uncensored"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "minas2025/Gemma-4-E2B-Coder-Uncensored",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/minas2025/Gemma-4-E2B-Coder-Uncensored

Ollama
How to use minas2025/Gemma-4-E2B-Coder-Uncensored with Ollama:
```
ollama run hf.co/minas2025/Gemma-4-E2B-Coder-Uncensored
```
Unsloth Desktop

How to use minas2025/Gemma-4-E2B-Coder-Uncensored with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf minas2025/Gemma-4-E2B-Coder-Uncensored

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": "minas2025/Gemma-4-E2B-Coder-Uncensored"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use minas2025/Gemma-4-E2B-Coder-Uncensored with Docker Model Runner:
```
docker model run hf.co/minas2025/Gemma-4-E2B-Coder-Uncensored
```

Lemonade

How to use minas2025/Gemma-4-E2B-Coder-Uncensored with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull minas2025/Gemma-4-E2B-Coder-Uncensored

Run and chat with the model

lemonade run user.Gemma-4-E2B-Coder-Uncensored-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use minas2025/Gemma-4-E2B-Coder-Uncensored with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf minas2025/Gemma-4-E2B-Coder-Uncensored

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 minas2025/Gemma-4-E2B-Coder-Uncensored

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use minas2025/Gemma-4-E2B-Coder-Uncensored with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf minas2025/Gemma-4-E2B-Coder-Uncensored

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 "minas2025/Gemma-4-E2B-Coder-Uncensored" \
  --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"

Coder-Gemma-4-E2B-Uncensored (Post-Mortem)

I tried to build a front-end coding specialist out of Gemma 4 E2B. It failed.

The base model is actually better at front-end development than this finetune is. It is also better at general Python. I am publishing this model and its full benchmark results anyway, because almost nobody publishes their failed runs, and the exact recipe for how I accidentally broke a good model might save you a few hours of compute.

If you just want a good small code model, go download the base llmfan46/gemma-4-E2B-it-ultra-uncensored-heretic. It beats this one on every metric.

The Autopsy (Benchmarks)

Everything below was measured with lm-evaluation-harness using greedy decoding and seed 0. Same harness, same conditions for both models.

General Python Logic

Model	HumanEval pass@1	MBPP pass@1
Base (heretic Q8_0)	66.5	43.5
This model (Q8_0)	56.7	40.5

Front-End Gate (12 React/TSX components)

Metric	Base (heretic)	This model
Structurally sound	12/12	12/12
Passes `tsc --noEmit` (strict)	10/12	9/12
Renders non-trivial HTML	11/12	10/12

The finetune lost about 10 points on abstract algorithmic logic, and it actually lost 1 point on the exact strict TypeScript compilation gate I trained it to improve.

Why it failed (Negative Transfer)

This is the part that is actually useful if you are training small models.

Gemma 4 E2B is already a highly optimized, very capable small model. When I injected about 700 synthetic front-end rows (Opus 4.6 and Claude Fable 5 distillates) and oversampled them to a third of the training updates, I did not teach it new front-end skills. Instead, I caused negative transfer.

The model learned to mimic the surface-level conversational style of the synthetic front-end data, but the weight updates scrambled the underlying logical routing that made the base model good at TypeScript in the first place. It traded actual reasoning for stylistic mimicry. The lesson here is that you cannot easily finetune a small, already-strong model to be "better" at a specific sub-skill using a tiny dataset of synthetic examples without degrading its general baseline.

Training Recipe (So you don't repeat it)

Property	Value
Base model	`llmfan46/gemma-4-E2B-it-ultra-uncensored-heretic`
Method	QLoRA 4-bit, merged to full model
LoRA rank / alpha	16 / 32
LoRA targets	q, k, v, o, gate, up, down, plus PLE gate and projection
Trainable params	~24M
Data	12,152 rows total (8k KodCode Python, ~700 unique front-end rows oversampled)
Training context	2048 tokens
Learning rate	1.5e-4, cosine schedule, 20 warmup steps
Batch	1 with grad accum 16 (effective 16)
Steps	260 (~47 min on a single RX 6700 XT)

Usage

If you still want to test it yourself, it is a standard Q8_0 GGUF.

llama.cpp:

llama-server -m Coder-Gemma-4-E2B-Uncensored.Q8_0.gguf -c 4096 -ngl 99 --jinja --port 8080

Downloads last month: 851

GGUF

Model size

5B params

Architecture

gemma4

Hardware compatibility

We're not able to determine the quantization variants.

View all variants

Model tree for minas2025/Gemma-4-E2B-Coder-Uncensored

Base model

google/gemma-4-E2B

Finetuned

google/gemma-4-E2B-it

Finetuned

llmfan46/gemma-4-E2B-it-ultra-uncensored-heretic

Quantized

(15)

this model

Datasets used to train minas2025/Gemma-4-E2B-Coder-Uncensored

Evaluation results

HumanEval pass@1 on openai_humaneval
self-reported

56.700
MBPP pass@1 on google-research-datasets/mbpp
self-reported

40.500