MiniMax-M2-REAP-162B-A10B-AWQ

MiniMax-M2-REAP-162B-A10B-AWQ

Base model: cerebras/MiniMax-M2-REAP-162B-A10B

Note: Some attention layers are left unquantized to preserve output coherence and consistency; 
as a result, the file size is reduced by about 24%, rather than the ~30% we might otherwise expect.

【Dependencies / Installation】

<i> Same as the original MiniMax-M2 </i>

As of 2025-11-19, create a fresh Python environment and run:

uv venv
source .venv/bin/activate
uv pip install 'triton-kernels @ git+https://github.com/triton-lang/[email protected]#subdirectory=python/triton_kernels' \
   vllm --extra-index-url https://wheels.vllm.ai/nightly --prerelease=allow

vLLM Official Guide

【vLLM Startup Command】

<i>Note: When launching with TP=8, include --enable-expert-parallel; otherwise the expert tensors wouldn’t be evenly sharded across GPU devices.</i>

CONTEXT_LENGTH=32768
vllm serve \
    QuantTrio/MiniMax-M2-REAP-162B-A10B-AWQ \
    --served-model-name MY_MODEL \
    --enable-auto-tool-choice \
    --tool-call-parser minimax_m2 \
    --reasoning-parser minimax_m2_append_think \
    --swap-space 8 \
    --max-num-seqs 32 \
    --max-model-len $CONTEXT_LENGTH \
    --gpu-memory-utilization 0.9 \
    --tensor-parallel-size 8 \
    --enable-expert-parallel \ 
    --trust-remote-code \
    --disable-log-requests \
    --host 0.0.0.0 \
    --port 8000

【Logs】

2025-11-19
1. Initial commit

【Model Files】

【Model Download】

from huggingface_hub import snapshot_download
snapshot_download('QuantTrio/MiniMax-M2-REAP-162B-A10B-AWQ', cache_dir="your_local_path")

【Overview】

<p align="center"> <em>𓌳 <strong>REAP</strong>𓌳 the Experts: Why Pruning Prevails for One-Shot MoE Compression</em><br> <img src="https://i.imgur.com/rmzG3gg.png" alt="REAP" width="75%"> </p>

MiniMax-M2-REAP-162B-A10B

✨ Highlights

Introducing MiniMax-M2-REAP-162B-A10B, a memory-efficient compressed variant of MiniMax-M2 that maintains near-identical performance while being 30% lighter.

This model was created using REAP (Router-weighted Expert Activation Pruning), a novel expert pruning method that selectively removes redundant experts while preserving the router's independent control over remaining experts. Key features include:

• Near-Lossless Performance: Maintains almost identical accuracy on code generation, agentic coding, and function calling tasks compared to the full 230B model
• 30% Memory Reduction: Compressed from 230B to 162B parameters, significantly lowering deployment costs and memory requirements
• Preserved Capabilities: Retains all core functionalities including code generation, math & reasoning and tool calling.
• Drop-in Compatibility: Works with vanilla vLLM - no source modifications or custom patches required
• Optimized for Real-World Use: Particularly effective for resource-constrained environments, local deployments, and academic research

---

📋 Model Overview

MiniMax-M2-REAP-162B-A10B has the following specifications:

• Base Model: MiniMax-M2
• Compression Method: REAP (Router-weighted Expert Activation Pruning)
• Compression Ratio: 30% expert pruning
• Type: Sparse Mixture-of-Experts (SMoE) Causal Language Model
• Number of Parameters: 162B total, 10B activated per token
• Number of Layers: 62
• Number of Attention Heads: 48
• Number of Experts: 180 (uniformly pruned from 256)
• Number of Activated Experts: 8 per token
• Context Length: 196,608 tokens
• License: Modified MIT

---

📊 Evaluations

<table> <thead> <tr> <th align="left">Benchmark</th> <th align="center">MiniMax-M2</th> <th align="center"><a href="https://huggingface.co/cerebras/MiniMax-M2-REAP-172B-A10B">MiniMax-M2-REAP-172B-A10B</a></th> <th align="center"><a href="https://huggingface.co/cerebras/MiniMax-M2-REAP-162B-A10B">MiniMax-M2-REAP-162B-A10B</a></th> <th align="center"><a href="https://huggingface.co/cerebras/MiniMax-M2-REAP-139B-A10B">MiniMax-M2-REAP-139B-A10B</a></th> </tr> </thead> <tbody> <tr> <td><strong>Compression</strong></td> <td align="center">—</td> <td align="center">25%</td> <td align="center">30%</td> <td align="center">40%</td> </tr> <tr> <td colspan="5" align="center"><strong>Coding</strong></td> </tr> <tr> <td><strong>HumanEval</strong></td> <td align="center">93.9</td> <td align="center">93.9</td> <td align="center">93.3</td> <td align="center">91.5</td> </tr> <tr> <td><strong>HumanEval+</strong></td> <td align="center">89.0</td> <td align="center">86.6</td> <td align="center">86.6</td> <td align="center">83.5</td> </tr> <tr> <td><strong>MBPP</strong></td> <td align="center">87.6</td> <td align="center">88.1</td> <td align="center">86.5</td> <td align="center">85.2</td> </tr> <tr> <td><strong>MBPP+</strong></td> <td align="center">73.0</td> <td align="center">74.9</td> <td align="center">73.0</td> <td align="center">71.4</td> </tr> <tr> <td colspan="5" align="center"><strong>Reasoning</strong></td> </tr> <tr> <td><strong>AIME25</strong></td> <td align="center">76.7</td> <td align="center">83.3</td> <td align="center">73.3</td> <td align="center">73.3</td> </tr> <tr> <td><strong>MATH-500</strong></td> <td align="center">91.6</td> <td align="center">89.4</td> <td align="center">89.4</td> <td align="center">93.8</td> </tr> <tr> <td colspan="5" align="center"><strong>Agentic / tool calling</strong></td> </tr> <tr> <td><strong>𝜏²-bench (Telecom, discard think traces)</strong></td> <td align="center">59.1</td> <td align="center">57.6</td> <td align="center">59.1</td> <td align="center">55.3</td> </tr> <tr> <td><strong>BFCLv3 (discard think traces)</strong></td> <td align="center">62.6</td> <td align="center">61.5</td> <td align="center">59.9</td> <td align="center">57.9</td> </tr> </tbody> </table>

🟩 This checkpoint maintains almost identical performance while being 30% lighter.

For more details on the evaluation setup, refer to the REAP arXiv preprint.

---

🚀 Deployment

You can deploy the model directly using the latest vLLM (that supports MiniMax-M2), no source modifications or custom patches required.

vllm serve cerebras/MiniMax-M2-REAP-162B-A10B \
    --tensor-parallel-size 8 \
    --tool-call-parser minimax_m2 \
    --reasoning-parser minimax_m2_append_think \
    --trust-remote-code \
    --enable_expert_parallel \
    --enable-auto-tool-choice

If you encounter insufficient memory when running this model, you might need to set a lower value for --max-num-seqs flag (e.g. set to 64). For more information, refer to the official vLLM deployment guide.

🧩 Model Creation

This checkpoint was created by applying the REAP (Router-weighted Expert Activation Pruning) method uniformly across all Mixture-of-Experts (MoE) blocks of MiniMax-M2, with a 30% pruning rate.

How REAP Works

REAP selects experts to prune based on a novel saliency criterion that considers both:

• Router gate values: How frequently and strongly the router activates each expert
• Expert activation norms: The magnitude of each expert's output contributions

This dual consideration ensures that experts contributing minimally to the layer's output are pruned, while preserving those that play critical roles in the model's computations.

Key Advantages

• One-Shot Compression: No fine-tuning required after pruning - the model is immediately ready for deployment
• Preserved Router Control: Unlike expert merging methods, REAP maintains the router's independent, input-dependent control over remaining experts, avoiding "functional subspace collapse"
• Generative Task Superiority: REAP significantly outperforms expert merging approaches on generative benchmarks (code generation, creative writing, mathematical reasoning) while maintaining competitive performance on discriminative tasks

📚 For more details, refer to the following resources:

• 🧾 arXiv Preprint
• 🧾 REAP Blog
• 💻 REAP Codebase (GitHub)

---

⚖️ License

This model is derived from MiniMaxAI/MiniMax-M2 and distributed under the modified MIT license.

---

🧾 Citation

If you use this checkpoint, please cite the REAP paper:

@article{lasby-reap,
  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}
}