MoA-Max: Best Mixture of Agents with models under 40B

A user proposes "MoA-Max," a local Mixture of Agents configuration using models under 40B parameters that achieves an overall score of 87.7, outperforming DeepSeek-v4-pro.

The setup utilizes Gemma-4-31B-QAT-Q4 and Ornith-1.0-35B as drafters with Qwen3.6-27B as the aggregator, requiring approximately 90GB of RAM.

• MoA-Max scores 89.2 on Bug Investigation and 91.8 on Multi-step Coding (DLQ).
• It ranks first among tested configurations, beating the second-place MoA variant by 0.9 points.
• The configuration outperforms standalone Qwen-3.6-27B (4-bit) which scored 86.6 overall.
• DeepSeek-v4-Pro achieved a lower average of 83.7 despite higher individual scores on some tasks.