OpenBMB Releases MiniCPM5-2B: A 2.52B Dense Model Averaging 53.9 Across 34 Benchmarks and Built to Run On Device

OpenBMB has released MiniCPM5-2B, the second checkpoint in its MiniCPM5 line and the follow-up to MiniCPM5-1B. It’s a dense causal language model with 2,516,756,480 parameters, 1,981,982,720 of them outside the embeddings, and it uses 42 layers with grouped-query attention. The context window is native, not bolted on: 131,072 tokens.

The practical pitch is simple. This is a standard LlamaForCausalLM model, so it can be loaded by mainstream engines without custom kernels or a model-code fork. OpenBMB also ships the weights under Apache 2.0, and says they run through vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS.

On the benchmark sheet, MiniCPM5-2B averages 53.9 across 34 rows. OpenBMB compares it with same-size models like LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it, while also listing bigger references such as Qwen3.5-4B, granite-4.2-3B, Nemotron-3-Nano-4B, Gemma-4-E4B-it and LFM2.5-8B-A1B. The best baseline in that table is Qwen3.5-4B at 51.1.

The model’s shape is pretty clear from the scores. It does well on code reasoning, with 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified, and it shows its biggest margin on tool use, including 97.1 on τ²-Bench Telecom and 66.6 on BFCL v4. It also leads on NoLiMa with 68.1, but it slips behind on some broader knowledge tests, including 70.8 on MMLU-Pro and 8.9 on Humanity’s Last Exam.

OpenBMB says the training stack runs from SFT to RL to on-policy distillation. The post-training phase starts with 400B tokens of deep-thinking SFT, then uses specialized RL teachers for math, code, agentic tasks and writing, and finishes with OPD, which merges 16 RL experts into one shipped model. The company says the open data release includes Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, UltraData-Math, UltraData-SFT-2605, UltraData-SFT-Agent-2609 with 500K agent samples, and UltraData-RL-2609 with more than 80K RL samples. It also published intermediate checkpoints, so the Base, Midtrain and SFT-only stages can be checked directly.