𓌳 REAP𓌳 the Experts: Why Pruning Prevails for One-Shot MoE Compression
📄 Paper • 💻 Code • 📝 Blog

MiniMax-M2.1-REAP-40

✨ Highlights

40% Expert-Pruned MiniMax-M2.1 optimized for code generation and function calling.

• 40% Expert Pruning: ~~96 experts remaining per layer
• Calibrated for Code & Tools: Same calibration mix as GLM-4.7 REAP models
• One-Shot Compression: No fine-tuning required

🙏 Acknowledgments

• Prime Intellect — Compute sponsorship (8x H200 cluster)
• Cerebras — REAP methodology

---

📋 Model Specifications

Property	Value
Base Model	MiniMax-M2.1
Compression	40% experts removed
Parameters	~264B
Experts per Layer	~96
Precision	BF16
Disk Size	~500GB

---

🔬 Calibration Dataset: Deep Dive

REAP's effectiveness depends critically on calibration data that represents the target use case. We specifically optimized for code generation, function/tool calling, and agentic workflows.

Why These 3 Datasets?

Dataset	Samples	Purpose	Why It Matters
evol-codealpaca-v1	700	Code generation	51% of mix — Code tasks activate specific expert pathways; pruning without code calibration destroys coding ability
xlam-function-calling-60k	330	Function/tool calling	24% of mix — Tool use requires structured JSON output; experts handling schema generation must be preserved
SWE-smith-trajectories	330	Agentic multi-turn	24% of mix — Real SWE-bench trajectories with tool calls, file edits, and multi-step reasoning

The Science Behind Dataset Selection

REAP Algorithm:
1. Forward pass calibration samples through model
2. Record which experts activate and their magnitudes
3. Compute saliency = router_weight × activation_norm
4. Prune lowest-saliency experts

Key Insight: Experts are TASK-SPECIFIC
├── Some experts specialize in natural language
├── Some experts specialize in code syntax
├── Some experts specialize in JSON/structured output
└── Some experts specialize in multi-turn context

If calibration lacks code → code-specialized experts appear "unused" → get pruned → model loses coding ability

Cerebras' Original Mix (from paper)

Cerebras used the same 3 datasets in their GLM-4.6 REAP experiments:

• evol-codealpaca-v1 for code generation
• xlam-function-calling-60k for tool calling
• SWE-smith-trajectories for agentic tasks

We followed this exact recipe for reproducibility.

Combined Dataset

Our calibration mix: 0xSero/glm47-reap-calibration-v2

---

📦 Related Models

Model	Compression	Experts	Size
MiniMax-M2.1-REAP-25	25%	~120	~620GB
MiniMax-M2.1-REAP-30	30%	~112	~580GB
MiniMax-M2.1-REAP-40	40%	~96	~500GB
MiniMax-M2.1-REAP-50	50%	~80	~420GB

---

🚀 Deployment

vLLM

vllm serve 0xSero/MiniMax-M2.1-REAP-40 \
    --tensor-parallel-size 8 \
    --trust-remote-code \
    --dtype bfloat16

---

🧩 Reproduction

REAP Pruning

#!/bin/bash
# MiniMax REAP - same calibration as GLM-4.7

export MODEL_DIR=/path/to/MiniMax-M2.1
export REAP_DATASET=0xSero/glm47-reap-calibration-mix
export REAP_SAMPLES_PER_CATEGORY=999
export REAP_MODEL_MAX_LENGTH=2048

python src/reap/prune.py \
    --model-name $MODEL_DIR \
    --dataset-name $REAP_DATASET \
    --compression-ratio 0.40 \
    --prune-method reap \
    --seed 42 \
    --distance_measure cosine

---

⚖️ License

Apache 2.0

---

🧾 Citation

@article{lasby2025reap,
  title={REAP the Experts: Why Pruning Prevails for One-Shot MoE Compression},
  author={Lasby, Mike and Lazarevich, Ivan and Sinnadurai, Nish and Lie, Sean and Ioannou, Yani and Thangarasa, Vithursan},
  journal={arXiv preprint arXiv:2510.13999},
  year={2025},
  url={https://arxiv.org/abs/2510.13999}
}