OpenBMB releases 2.52B MiniCPM5-2B language model
OpenBMB has released MiniCPM5-2B, a dense causal language model with 2,516,756,480 parameters and a native context window of 131,072 tokens. It uses a standard LlamaForCausalLM architecture, 42 layers and grouped-query attention with 16 query heads and two key/value heads. The model is available under Apache 2.0 and is supported by vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS.
Across 34 benchmark rows, MiniCPM5-2B records an average score of 53.9, compared with 51.1 for Qwen3.5-4B. Its largest reported advantages are in tool use and coding: it scores 97.1 on τ²-Bench Telecom, 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified. Results are weaker on some general-knowledge tests, including MMLU-Pro and Humanity’s Last Exam, and mixed on long-context evaluations.
Training combined supervised fine-tuning, reinforcement learning and on-policy distillation. OpenBMB says the post-training process used 400B tokens for deep-thinking SFT, specialised RL teachers and 16 merged RL experts. The company also released training datasets and intermediate Base, Midtrain and SFT-only checkpoints for analysis.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.