ModelBest Releases MiniCPM5-2B With Training Stack for Edge Devices

ModelBest and the OpenBMB community have released MiniCPM5-2B, an open-weights language model aimed at edge devices. Rather than publishing only a downloadable model, the release includes datasets, training recipes and reinforcement-learning infrastructure alongside the weights.

More than a model checkpoint

Open weights are the model parameters developers can download and run themselves. Artificial Analysis lists MiniCPM5-2B as open weights under the Apache 2.0 license. ModelBest says its package also provides the materials used around the model’s development, including its training recipes and reinforcement-learning infrastructure.

That changes what developers receive. A weights-only release supplies a system to run or fine-tune; this package also supplies training inputs and methods. ModelBest presents those materials as a route for researchers and developers to fine-tune specialized edge applications, though the release does not establish how readily its results can be reproduced on a particular device.

14MiniCPM5-2B Intelligence Index

Artificial Analysis gives MiniCPM5-2B an estimated Intelligence Index score of 14.

6Comparable-model median

Artificial Analysis lists a median score of 6 among comparable models.

A text model for local tasks

MiniCPM5-2B accepts and produces text, with a 131,000-token context window. Artificial Analysis classifies it as a reasoning model and lists a knowledge cutoff of December 31, 2025. It is not a multimodal system: the profile says it cannot process images.

ModelBest says the model supports tool calling, deep search, code generation and multi-step reasoning. It names PCs, smartphones, robotics and IoT hardware as deployment targets, with document processing, data synthesis and multi-turn question answering among the intended local uses.

The evaluation is not final

Artificial Analysis labels the 14-point result an estimate and says independent evaluation is forthcoming. Its Intelligence Index v4.3 combines 10 evaluations, including tests for agentic work, coding, reasoning, document reasoning and long-context tasks. ModelBest separately says MiniCPM5-2B ranked first on the index among open-source models below 4 billion parameters.

The sources use different size figures: ModelBest calls it a 2-billion-parameter model, while Artificial Analysis lists 2.6 billion total parameters. Artificial Analysis also lists zero pricing for one million input and output tokens. Those entries do not measure the hardware, power or operating costs of self-hosting on any particular device.