MiniCPM5-1B is a 1.08 billion parameter dense causal language model developed by OpenBMB and released in May 2026. Designed primarily for on-device and resource-constrained environments, it serves as the initial checkpoint in the MiniCPM5 series. The model aims to provide high performance in coding, tool-calling, and reasoning tasks while maintaining a small memory footprint suitable for local deployment.
The model utilizes a standard LlamaForCausalLM architecture, featuring 24 layers and Grouped-Query Attention (GQA) with 16 query heads and 2 key-value heads. It is distinguished by a context window of 131,072 tokens, which is exceptionally large for its size class. This allows the model to process extensive documents and manage complex multi-step agentic workflows that require significant local memory.
MiniCPM5-1B features a Hybrid Reasoning capability within a single checkpoint. Using a specific chat template, the model can operate in either a standard assistant mode or a deliberate reasoning mode. In the "non-reasoning" or standard mode, the model functions as an efficient instruction-following assistant, allowing users to toggle behaviors based on task complexity without needing to load separate model weights.
The training process utilized the UltraData tiered management framework, which involved stable base training, mid-training for capability adaptation, and a multi-step post-training phase. Post-training included Supervised Fine-Tuning (SFT), Reinforcement Learning (RL), and On-Policy Distillation (OPD). OPD was employed to distill reasoning strategies from larger teacher models, enhancing the 1B model’s performance in math, logic, and coding.
Released under the Apache 2.0 license, the model is part of a broader ecosystem including the Ultra-FineWeb and UltraData-SFT datasets. Evaluations indicate that it is particularly effective at agentic tool use and code generation, often outperforming larger models in the sub-2B parameter category on intelligence and efficiency benchmarks.