MiniCPM5-1B is a dense 1.1 billion parameter causal language model developed by OpenBMB, representing the first release in the MiniCPM5 series. Designed primarily for on-device and edge-side deployment, it is optimized to run in resource-constrained environments while providing performance competitive with larger open-weights models. The model is built on a standard Llama architecture, featuring 24 layers and Grouped-Query Attention (GQA), ensuring compatibility with most mainstream inference backends.
A defining feature of the model is its hybrid reasoning capability, which allows a single checkpoint to function as either a fast assistant or a deliberate reasoner. This is managed through a built-in chat template that supports a <think> block, which users can toggle by enabling or disabling the thinking mode (via the enable_thinking parameter in supported templates). When reasoning is enabled, the model performs intensive chain-of-thought processing, significantly improving its performance in complex tasks such as competition-level mathematics and logic.
In terms of capabilities, MiniCPM5-1B shows particular strength in agentic tool use, code generation, and mathematical reasoning. It has demonstrated high accuracy on benchmarks like AIME 2026 and MATH-500, often surpassing models with significantly higher parameter counts. The model also supports a native long-context window of 131,072 tokens (128K), making it suitable for processing extensive documents or long conversation histories locally.
For optimal results in reasoning mode, the model requires a higher max_tokens setting (typically 32K to 64K) to allow the full expression of its internal reasoning chains. OpenBMB provides specific "Agent Skills" and cookbooks to facilitate fine-tuning and deployment for specialized tasks such as local assistants and coding agents.