OpenBMB Releases MiniCPM5-2B Model for Edge Devices

OpenBMB has officially released MiniCPM5-2B, a dense artificial intelligence model featuring 2.52 billion parameters. The model weights are distributed under an Apache 2.0 license and are designed to run smoothly through inference frameworks such as vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX, and FlagOS, enabling efficient execution directly on local hardware and edge devices.

In benchmark evaluations, OpenBMB compared MiniCPM5-2B against similar-sized models including LFM2.5-2.6B, Qwen3.5-2B, and Gemma-4-E2B-it, alongside reference models like Qwen3.5-4B, granite-4.2-3B, Nemotron-3-Nano-4B, Gemma-4-E4B-it, and LFM2.5-8B-A1B. Across 34 evaluation benchmarks, MiniCPM5-2B achieved an overall average score of 53.9, outperforming other baseline models in its specific size class.

Looking into specific capabilities, the model scores 69.1 on LiveCodeBench v6 for code reasoning and 46.4 on SWE-bench Verified. Tool use represents its widest performance margin, recording 97.1 on τ²-Bench Telecom, 66.6 on BFCL v4, and 20.8 on τ³-Bench Banking. Long-context performance reaches 68.1 on NoLiMa, while general knowledge remains constrained by its size, posting 70.8 on MMLU-Pro and 8.9 on Humanity’s Last Exam.

Stock photo for illustration only, not from the actual event

The training pipeline follows the UltraData tiered data management method outlined in original research, passing through stable and decay base training phases followed by mid-training adaptation. Post-training incorporates 400 billion tokens of deep-thinking SFT and specialized RL teachers trained via the JustRL II algorithm for math, code, agents, and writing. The final phase applies on-policy distillation (OPD), merging 16 RL experts into a single model to boost reasoning and agentic benchmark scores.