Cerebras Releases MiniMax-M2-REAP-162B-A10B: A Memory Efficient Version of MiniMax-M2 for Long Context Coding Agents

Cerebras Releases MiniMax-M2-REAP-162B-A10B: A Memory Efficient Version of MiniMax-M2 for Long Context Coding Agents

Cerebras introduced MiniMax-M2-REAP-162B-A10B, a Sparse Mixture-of-Experts (SMoE) model derived from MiniMax-M2, achieving 30% expert pruning while retaining 10B active parameters per token. The model maintains performance on coding and reasoning benchmarks despite reducing total parameters by 30%.

Why This Matters

Large SMoE models like MiniMax-M2 (230B total parameters) are computationally heavy for deployment. Traditional expert merging risks “functional subspace collapse,” degrading performance. REAP pruning avoids this by selectively removing low-saliency experts, preserving router control and achieving near-lossless compression at 30% pruning, as shown on HumanEval (90% accuracy) and MBPP (80% accuracy).

Key Insights

• “30% expert pruning, 2025”: Cerebras’ REAP method reduces MiniMax-M2 from 230B to 162B parameters while retaining 10B active per token.
• “Sagas over ACID for e-commerce”: Not applicable here; REAP’s pruning outperforms expert merging for generative tasks.
• “vLLM used by Cerebras”: Deployment example shows vllm serve with --tensor-parallel-size 8 for efficient inference.

Working Example

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

Practical Applications

• Use Case: Coding agents using long-context LLMs (e.g., HumanEval, MBPP).
• Pitfall: Over-pruning beyond 30% may degrade performance on mathematical reasoning (AIME 25, MATH 500).

References:

• https://www.marktechpost.com/2025/11/15/cerebras-releases-minimax-m2-reap-162b-a10b-a-memory-efficient-version-of-minimax-m2-for-long-context-coding-agents/

---