MiniCPM5-1B, a 1-billion-parameter model from OpenBMB, scores an average of 42.57 on agentic and reasoning benchmarks, beating the next-best 1B-class competitor's 35.61. It supports native tool calling and the Model Context Protocol (MCP) out of the box, enabling local agent workflows on consumer hardware without cloud connectivity. The model uses InfLLM v2 attention mechanism to process long contexts efficiently, with a 128K token window (roughly 96,000 words). Post-training via reinforcement learning and knowledge distillation boosted benchmark scores by 16 points. In tests, MiniCPM5-1B showed strong conversational fluency but failed a classic logic trap (marrying widow's sister) and hedged on an A/B choice question. However, when paired with an MCP server, agentic tasks like providing Bitcoin price and stock recommendations worked well. The model is available on Hugging Face under Apache 2.0 license.

Key facts

• MiniCPM5-1B averages 42.57 on benchmarks, outperforming other 1B models.
• Supports native tool calling and MCP for offline agent workflows.
• 128K token context window enables long conversations and document processing.
• Failed a classic logic trap and showed hedging in A/B choice tests.
• Available on Hugging Face under Apache 2.0 license.

KeyAudit data perspective