MiniCPM

Why we included this project

Developers who need a genuinely capable language model on constrained hardware, from phones to laptops, are the main audience here. MiniCPM5-1B is a dense 1B-parameter Transformer that posts an average benchmark score well above other strong open models of the same size, with the biggest gains in agentic tool use, code, and competition math. The same checkpoint can act as a quick assistant or a slower, more deliberate reasoner depending on a built-in thinking toggle you switch per request, so you don't need to run two separate models. The repo also ships single-page cookbooks and Agent Skills for the common inference backends and fine-tuning frameworks, which shortens the jump from a demo to something you can actually deploy. A bundled desktop-pet example runs the model locally and is a lightweight way to see it in action before you wire it into your own product.