TL;DR
MiniCPM5-1B is a 1B-class on-device LLM that achieves SOTA with an average score of 42.57 compared to peers, and includes built-in hybrid reasoning and agent skills.
Key features
Top performance: Records the highest average score (42.57) among 1B-class open-source models, excelling particularly in agent tool use, code, and competitive math.
Hybrid reasoning: Built-in <think> template with enable_thinking switch allows switching between fast responses and deep reasoning.
Deployment/fine-tuning agent skills: Provides single-page recipes and Agent Skills for major inference backends and fine-tuning frameworks.
Desktop pet: Includes a local LLM-based desktop pet application.
When to use it
When a powerful 1B-class LLM is needed for on-device, local deployment, or resource-constrained environments.
When leveraging hybrid reasoning (fast response + deep thinking) from a single checkpoint.
When quickly deploying or fine-tuning a lightweight model with agent capabilities (tool use, code execution, etc.).