Qwen1.5-MoE-A2.7B is a small mixture-of-expert (MoE) model with only 2.7 billion activated parameters yet matches the performance of state-of-the-art 7B models like Mistral 7B and Qwen1.5-7B.