According to CoinMeta, Kimi launched the K2.5 version of agent swarm in January this year, which allows a main agent to recruit and command up to 300 sub-agent. Its reinforcement learning method PARL trains only the "boss" responsible for command, while freezing the sub-agent. The General Reasoning team introduced a 120B parameter model Avalon-M1-Mini and proposed collective reward reinforcement learning (RLCR), where one agent acts as the CEO, recruiting up to 8 employees, and collaborating through message boards and shared workspaces. This method has seen continuous improvement in 147 training environments, with the proportion of lead engineers in recruitment roles rising from 0.4% to about 13%. Currently, Avalon-M1-Mini is still in training, and the team plans to publish the final paper upon completion of training, as well as open-source the training environment, agent framework, and model weights.