Instructions to use kueizen/Marco-Mini-Instruct-REAP-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 kueizen/Marco-Mini-Instruct-REAP-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 kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf kueizen/Marco-Mini-Instruct-REAP-GGUF: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 kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M
# Run inference directly in the terminal:
./llama-cli -hf kueizen/Marco-Mini-Instruct-REAP-GGUF: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 kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama-cli -hf kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

Use Docker

docker model run hf.co/kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

LM Studio
Jan

vLLM

How to use kueizen/Marco-Mini-Instruct-REAP-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "kueizen/Marco-Mini-Instruct-REAP-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": "kueizen/Marco-Mini-Instruct-REAP-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

Ollama
How to use kueizen/Marco-Mini-Instruct-REAP-GGUF with Ollama:
```
ollama run hf.co/kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M
```
Unsloth Desktop

How to use kueizen/Marco-Mini-Instruct-REAP-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

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": "kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use kueizen/Marco-Mini-Instruct-REAP-GGUF with Docker Model Runner:
```
docker model run hf.co/kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M
```

Lemonade

How to use kueizen/Marco-Mini-Instruct-REAP-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

Run and chat with the model

lemonade run user.Marco-Mini-Instruct-REAP-GGUF-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use kueizen/Marco-Mini-Instruct-REAP-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 kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

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 kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use kueizen/Marco-Mini-Instruct-REAP-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M

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 "kueizen/Marco-Mini-Instruct-REAP-GGUF:Q4_K_M" \
  --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"

Marco-Mini-Instruct · REAP expert-pruned GGUFs

Marco-Mini-Instruct (17.3B total, ~0.86B active, 256 experts, 8 active per token), with 5% to 70% of its experts removed by REAP. Nine files from 9.3 GB down to 3.1 GB, all Q4_K_M, all measured on one RTX 3090.

Files

File	Experts kept	Size (GB)	PPL	pp512 (t/s)	tg128 (t/s)
unpruned base (not in this repo)	256	10.00	8.17	6,221	—
`Marco-Mini-Instruct-REAP05-Q4_K_M.gguf`	244	9.33	8.59	5,160	155¹
`Marco-Mini-Instruct-REAP10-Q4_K_M.gguf`	231	8.85	9.30	6,482	281
`Marco-Mini-Instruct-REAP15-Q4_K_M.gguf`	218	8.36	10.28	6,806	256
`Marco-Mini-Instruct-REAP20-Q4_K_M.gguf`	205	7.88	11.40	6,005	277
`Marco-Mini-Instruct-REAP30-Q4_K_M.gguf`	179	6.94	14.64	7,661	278
`Marco-Mini-Instruct-REAP40-Q4_K_M.gguf`	154	5.97	18.74	6,970	278
`Marco-Mini-Instruct-REAP50-Q4_K_M.gguf`	128	5.00	26.42	6,966	316
`Marco-Mini-Instruct-REAP60-Q4_K_M.gguf`	102	4.07	38.62	9,335	287
`Marco-Mini-Instruct-REAP70-Q4_K_M.gguf`	77	3.10	64.43	9,713	278

¹ Measured in a separate session from the other rows; treat as an outlier, not a slowdown.

KL divergence is not listed for this family yet: our two measurement sessions disagree with each other, and we'd rather show nothing than a number we can't stand behind.

Which one to pick

REAP10–REAP15 (8.9–8.4 GB): the sweet spot, at most 1.26× the base perplexity.
REAP30 (6.9 GB): the best balance of size and quality if you need room on a 24 GB card for long context or a second model.
Under 5 GB: don't prune Mini further. Use Marco-Nano-Instruct-REAP-GGUF instead (see below).

The first 10% is nearly free (1.14× base perplexity). Past that, Mini loses quality faster than Nano: at REAP30 it sits at 1.79× its base, where Nano sits at 1.57×.

Generation speed barely changes with pruning, because a MoE model runs 8 experts per token whatever the total. Pruning buys you size, not speed.

Pick down, don't prune down

Across our sweeps, a smaller model pruned lightly beats a bigger model pruned hard at the same file size. Both Marco models share a tokenizer (151,936 tokens), so their perplexities compare directly:

Size budget	Marco-Mini, pruned hard	Marco-Nano, pruned lightly
~5 GB	REAP50 · 5.00 GB · PPL 26.42	REAP05 · 4.77 GB · PPL 12.07
~4 GB	REAP60 · 4.07 GB · PPL 38.62	REAP20 · 4.05 GB · PPL 15.36
~3 GB	REAP70 · 3.10 GB · PPL 64.43	REAP40 · 3.11 GB · PPL 22.71

Rule of thumb: choose the smallest model that fits your memory, then prune it 5–20%. Past about 30%, each extra step costs a lot more quality than the one before.

How these were made

Pruning: REAP (Router-weighted Expert Activation Pruning, Cerebras Research), using the reference implementation. REAP scores each expert by its router weight times the size of its output over a calibration set, then removes the lowest-scoring experts whole. It is one-shot: no retraining. Seed 42.
Calibration set: theblackcat102/evol-codealpaca-v1 (train split, shuffled, seed 42), 64 samples per category, batch size 1, max sequence length 2048 tokens.
Quantisation: Q4_K_M with llama.cpp.
REAPxx in a file name is the share of experts removed, e.g. REAP20 = 20% of experts removed.

How we measured

Perplexity (PPL): WikiText-2 raw test split, 50 chunks, context 512, llama-perplexity. Lower is better.
Speed: llama-bench on one RTX 3090 (24 GB), all layers on GPU: prompt processing over 512 tokens (pp512) and generation over 128 tokens (tg128), 3 runs, llama.cpp builds b8164 and b8575.
Size: file size in GB.

Perplexity tracks how close a pruned model stays to its base, but it is not a task benchmark. We have not run MMLU, coding or multilingual evals on these files, so test them on your own workload before relying on them.

Run it

llama.cpp (OpenAI-compatible server):

llama-server --hf-repo kueizen/Marco-Mini-Instruct-REAP-GGUF --hf-file Marco-Mini-Instruct-REAP15-Q4_K_M.gguf -ngl 99

Ollama: this repo holds several files with the same quant type, so download the one you want and point a Modelfile at it:

FROM ./Marco-Mini-Instruct-REAP15-Q4_K_M.gguf
PARAMETER num_ctx 8192

ollama create marco-mini-instruct-reap15 -f Modelfile
ollama run marco-mini-instruct-reap15

LM Studio: search for kueizen/Marco-Mini-Instruct-REAP-GGUF and pick a file.

License and credits

These files are derived from ATH-MaaS/Marco-Mini-Instruct and released under the same Apache 2.0 licence. What we changed: removed experts with REAP and quantised the result to GGUF Q4_K_M. Nothing else was modified or retrained.

REAP is by Cerebras Research: paper, code. All credit for the base model goes to its authors.

The series

Published by Kueizen.

Downloads last month: 254

GGUF

Model size

16B params

Architecture

qwen3moe

Hardware compatibility

4-bit

Model tree for kueizen/Marco-Mini-Instruct-REAP-GGUF

Base model

ATH-MaaS/Marco-Mini-Instruct

Quantized

(8)

this model

Paper for kueizen/Marco-Mini-Instruct-REAP-GGUF

REAP the Experts: Why Pruning Prevails for One-Shot MoE compression

Paper • 2510.13999 • Published Oct 15, 2025 • 23