OpenBMB just dropped MiniCPM5-2B, and the numbers suggest a shift in how we evaluate small language models. This isn’t just another parameter count bump; it is a dense causal model with exactly 2,516,756,480 parameters that prioritizes agent behavior over raw trivia knowledge. By sticking to the standard LlamaForCausalLM architecture, OpenBMB ensured immediate compatibility with vLLM, Ollama, and llama.cpp. No custom kernels, no forked code. It just works.
The benchmark results tell a specific story. Across 34 tests, MiniCPM5-2B averaged 53.9, outperforming the larger Qwen3.5-4B, which scored 51.1. But the real divergence appears in tool use. On the τ²-Bench Telecom test, MiniCPM5-2B hit 97.1, leaving competitors far behind. It also scored 69.1 on LiveCodeBench v6. However, do not expect it to ace general knowledge exams. It trailed Qwen on MMLU-Pro (70.8 vs 78.0) and Humanity’s Last Exam. This model is built to do things, not just recite facts.
How did they get there? The training recipe is unusually transparent. OpenBMB started with 400 billion tokens of deep-thinking supervised fine-tuning. Then came the critical step: on-policy distillation (OPD). Instead of building a new corpus, they merged 16 specialized RL experts—five focused on agentic tasks—into a single student model. They used reverse KL divergence to align the student with teacher logits. OpenBMB claims this stage alone added nearly 11 points to reasoning benchmarks.
Perhaps the most significant move is the data release. OpenBMB didn’t just share weights; they published Ultra-FineWeb, UltraData-Code, and over 80,000 RL samples. Intermediate checkpoints for Base, Midtrain, and SFT stages are also available. This level of openness allows the community to verify the impact of each training phase. For developers building local agents, MiniCPM5-2B offers a rare combination of deployability and verifiable performance. It proves that specialized distillation can outperform brute-force scaling in narrow, high-value domains.
