On September 8, 2026, Minimax, in collaboration with OpenBMB, officially announced the open-source release of MiniCPM5-2B, a cutting-edge language foundation model tailored for end-user devices. Despite its compact size of only 2 billion (2B) parameters, this model boasts a robust suite of functionalities, including tool invocation, in-depth search capabilities, and code generation. These features collectively mark a significant stride towards realizing the vision of a general-purpose AI agent operating seamlessly on end-user devices.

In the competitive landscape of AI models, MiniCPM5-2B has distinguished itself by topping the Artificial Analysis (AA) leaderboard among open-source models with fewer than 4B parameters, securing an impressive score of 23. Notably, it has also achieved a commanding lead in agent capability evaluation, outperforming competitors with a score of 20. This accomplishment underscores the model's advanced capabilities and potential for real-world applications.

To ensure broad compatibility and ease of integration, MiniCPM5-2B has been meticulously optimized for mainstream chips from industry leaders such as AMD, Intel, MediaTek, and Qualcomm. Furthermore, in a collaborative effort with the FlagOS community, the model has undergone Day 0 adaptation for nine additional chips, including Huawei Ascend and NVIDIA, ensuring seamless performance across a diverse range of hardware platforms.

One of the model's most compelling attributes is its efficiency. Following INT4 quantization, the model's size is reduced to a mere 1.2GB, enabling it to run smoothly on devices with as little as 4GB of memory. This makes MiniCPM5-2B an ideal candidate for developing localized AI assistants on a variety of end-user devices, including smartphones, PCs, and smart cockpits, thereby democratizing access to sophisticated AI capabilities.