On September 8, Langboat AI, in collaboration with OpenBMB, officially released the open-source code for MiniCPM5-2B, a cutting-edge on-device language model. Boasting 2 billion parameters, this model soared to the top of the global open-source models with parameters under 4 billion on the Artificial Analysis Intelligence Index, achieving a remarkable score of 23. Additionally, it led its class in the Agentic Index evaluation with a score of 20, showcasing its preliminary capabilities as a general agent on edge devices. These capabilities include support for tool invocation, in-depth search functionalities, and code generation.

Accompanying the model, Langboat AI also open-sourced its efficient data governance system, UltraData, along with its self-developed reinforcement learning frameworks, Meshy and JustRL II. These components are designed to be Day0-compatible with chips from leading manufacturers such as Intel, Rockchip, and Arm. This compatibility significantly slashes both inference costs and latency on edge devices, paving the way for accelerated commercial deployment of on-device AI solutions.