# MiniCPM

A family of small, on-device large language models for edge and resource-constrained deployment, with the newest 2-billion-parameter release reporting state-of-the-art results for its size class.

MiniCPM is a family of small, on-device large language models built for edge and resource-constrained deployment. The newest release, MiniCPM5-2B, is a dense 2-billion-parameter transformer that scales up the training recipe used for the earlier MiniCPM5-1B model, and the project reports it reaches state-of-the-art results among open-source models of similar size while staying competitive with some 4-billion-parameter models, particularly in coding, mathematics, long-context understanding, tool use and agentic tasks.

Alongside the model, the project publishes the training datasets behind it under an 'UltraData' umbrella: a web pretraining dataset, a tiered code dataset used to improve coding ability, a set of 500,000 agent-training samples aimed at on-device agent capability, and a further set of more than 80,000 reinforcement-learning samples covering mathematics, code, general knowledge and long-context reasoning.

Earlier releases in the same family include a sparse-and-linear hybrid-attention model built for million-token context, a sparse-attention model with hybrid reasoning, and an edge-focused model reporting substantially faster generation on typical edge chips.

Website: https://github.com/OpenBMB/MiniCPM