1,080,632,832
131,072 tokens
Apache 2.0
Official open-weight MiniCPM5 family release covering the final MiniCPM5-1B checkpoint and the same-day SFT-only checkpoint that precedes RL and OPD post-training. OpenBMB positions MiniCPM5-1B as a dense 1B-class LlamaForCausalLM model for on-device and resource-constrained deployment, with native 131,072-token context, 24 layers, 16 query heads with 2 KV heads under GQA, built-in think and no-think chat modes through the same checkpoint, and native XML-style tool-calling support with SGLang parser integration. The release card claims 1B-class open-source SOTA within its comparison set, with especially visible gains in agentic tool use, code generation, and difficult reasoning versus small open models such as Qwen3 and Qwen3.5 think variants and LFM2.5-1.2B-Thinking. Training materials describe a three-stage stack of base training, mid-training, and post-training under OpenBMB's UltraData tiered data management framework, using Ultra-FineWeb, Ultra-FineWeb-L3, UltraData-Math, and UltraData-SFT-2605. For the final checkpoint, the card highlights a post-training pipeline of SFT, specialized RL teachers, and On-Policy Distillation, with claimed average gains of 16 points on math, code, and instruction-following tasks while reducing max-token-budget overlong responses by 29 percentage points. The grouped SFT checkpoint represents the pre-RL and pre-OPD stage, while the final BF16 release is the fully post-trained model intended for local assistants, coding agents, tool-use workflows, and compact reasoning assistants.