OpenBMB published the weights of MiniCPM5-2B on September 7, 2026, a 2.52 billion-parameter language model under an Apache 2.0 license that scores an average of 53.9 points across the 34 benchmarks the lab picked, against 51.1 for Qwen3.5-4B, according to the authors. The 2B is the second chapter of the MiniCPM5 series, after the one-billion-parameter model released on May 19, and carries the signature of OpenBMB, the open source community tied to Tsinghua University and ModelBest. The OpenBMB/MiniCPM repository boasts 11,131 stars on GitHub as of September 20, 2026, and according to Tencent News the whole MiniCPM family passed 50 million cumulative downloads at the end of August. MiniCPM5-2B is a dense, text-only model with a 131,072-token context window, a standard LlamaForCausalLM architecture with 42 layers and grouped-query attention (16 query heads against 2 for keys and values), and a single checkpoint offering two modes switched via the enable_thinking parameter. Training runs in three stages, with the third combining SFT with reinforcement learning and on-policy distillation; OpenBMB states reinforcement and distillation added an average of 10.96 points on reasoning and 6.96 on agent