Rushd-Agent

An efficient MoE-based agentic system with NCI-driven early exit for Arabic natural language understanding.

Architecture

• Router: Qwen2.5-7B with quantile-balanced expert selection
• Experts: 6 domain-specialized Qwen3.5 models (Geo 27B + Writer/Game/Decision/Plan/Logic 9B)
• Early Exit: Noema Predictor (21M params, 98.3% accuracy, 3.8x speedup)
• API: OpenAI-compatible server on port 8777

Paper

See rushd_agent_paper.pdf for full details.

Citation

@misc{yahya2025rushd,
  title={Rushd-Agent}: An Efficient MoE-Based Agentic System with NCI-Driven Early Exit,
  author={Yahya},
  year={2025},
  publisher={Hugging Face}
}