---
license: apache-2.0
base_model: ATH-MaaS/Marco-Mini-Instruct
library_name: transformers
pipeline_tag: text-generation
tags:
- moe
- mixture-of-experts
- reap
- expert-pruning
---

# Marco-Mini-Instruct-REAP40

[Marco-Mini-Instruct](https://huggingface.co/ATH-MaaS/Marco-Mini-Instruct) (17.3B total, ~0.86B active, 256 experts, 8 active per token) with **40% of its experts removed** by [REAP](https://arxiv.org/abs/2510.13999): 154 of 256 experts kept. These are the full-precision (bf16) safetensors weights, for fine-tuning, re-quantising or running with `transformers`.

For ready-to-run quantised files at every pruning ratio, see [kueizen/Marco-Mini-Instruct-REAP-GGUF](https://huggingface.co/kueizen/Marco-Mini-Instruct-REAP-GGUF).

## Load it

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "kueizen/Marco-Mini-Instruct-REAP40"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(repo, torch_dtype="auto", device_map="auto")
```

## How this was made

- **Pruning:** [REAP](https://arxiv.org/abs/2510.13999) (Router-weighted Expert Activation Pruning, Cerebras Research), using the [reference implementation](https://github.com/CerebrasResearch/reap). 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](https://huggingface.co/datasets/theblackcat102/evol-codealpaca-v1) (train split, shuffled, seed 42), 64 samples per category, batch size 1, max sequence length 2048 tokens.
- **Experts removed:** 40% (154 of 256 kept).

We have not run task benchmarks on these checkpoints (MMLU, coding, multilingual). Test them on your own workload before relying on them.

## License and credits

Derived from [ATH-MaaS/Marco-Mini-Instruct](https://huggingface.co/ATH-MaaS/Marco-Mini-Instruct) and released under the same **Apache 2.0** licence. What we changed: removed experts with REAP. Nothing else was modified or retrained. REAP is by Cerebras Research: [paper](https://arxiv.org/abs/2510.13999), [code](https://github.com/CerebrasResearch/reap). All credit for the base model goes to its authors.

Published by [Kueizen](https://kueizen.com).