Instructions to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-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 EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-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 EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF
# Run inference directly in the terminal:
llama cli -hf EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF
# Run inference directly in the terminal:
llama cli -hf EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

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 EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF
# Run inference directly in the terminal:
./llama-cli -hf EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

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 EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF
# Run inference directly in the terminal:
./build/bin/llama-cli -hf EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

Use Docker

docker model run hf.co/EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

LM Studio
Jan

vLLM

How to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

Ollama
How to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF with Ollama:
```
ollama run hf.co/EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF
```
Unsloth Desktop

How to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

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": "EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF with Docker Model Runner:
```
docker model run hf.co/EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF
```

Lemonade

How to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

Run and chat with the model

lemonade run user.gemma-4-26B-A4B-it-heretic-APEX-GGUF-{{QUANT_TAG}}

List all available models

lemonade list

Hermes Agent

How to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-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 EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

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 EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF

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 "EmanuelOverride/gemma-4-26B-A4B-it-heretic-APEX-GGUF" \
  --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"

Praise to HarmonicMinus

by Koshkasa - opened 25 days ago

Discussion

Koshkasa

25 days ago

•

edited 25 days ago

In testing, HarmonicMinus appears to be the sweet(est) spot for a 16-24GB GPU, although it takes a fully vacant 16GB GPU, small context, or aggressive KV cache quantization. That said, with KV cache in RAM, it correctly utilizes tool calls and stays on track way past 60k - I suppose I don't have a sufficiently large NIHS set to test it against.

To claw back some VRAM space on sm_120 capable GPUs specifically, what I tried is an edit of HarmonicMinus. It involves capping ffn_down and ffn_up at Q6_K instead of Q8_0, and every sparse expert tensor that was either at Q4_K, or ffn_down_exps Q3_K falling back to Q4_0 during quantization, now switched to MXFP4. The result is a 4.48bpw quant, 13.1GB in size. I am currently testing it against HarmonicMinus, but so far, the results seem promising.

If file size isn't a major concern, the regained extra 0.3 bpw budget could be spent on quantizing global attn to q8_0 or slightly increasing quality layer allocation.

My main conclusion from trying to utilize MXFP4 in APEX is never to quant attention at MXFP4, under any circumstance - the coarse attn outlier precision hits Gemma 4 particularly hard, breaks tool calls, punctuation, and on rare occasions, grammar too. The lower precision ffn_down_exps that were quantized to q4_0 seem not to suffer hard from lower outlier precision in mxfp4

UPD: recipe and test quants for comparison uploaded here.

EmanuelOverride

Owner 22 days ago

•

edited 22 days ago

Cheers, great idea and I'd love to test your MXFP4 variant. Swamped with life at the moment, but not abandoning the project, just forced to take a little break.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment