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 the MiniCPM5 series after MiniCPM5-1B. It is a dense causal model with about 2.52 billion parameters (roughly 1.98B outside embeddings), 42 layers, grouped-query attention (16 query heads, 2 key/value heads), and a native 131,072-token context. The architecture is standard LlamaForCausalLM, so vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX, and FlagOS can load it without custom kernels.

Across 34 benchmark rows it averages 53.9, ahead of Qwen3.5-4B at 51.1 in the reported table, with granite-4.2-3B at 42.7 and LFM2.5-2.6B at 33.2. Weights are Apache 2.0, and OpenBMB is also releasing pre-training, SFT, and RL datasets plus intermediate Base, Midtrain, and SFT-only checkpoints—rare transparency for a claim built on RL plus on-policy distillation.

The scorecard is specialized, not generalist. On coding agents, MiniCPM5-2B posts 69.1 on LiveCodeBench v6 versus 56.4 for the compared baseline in MarkTechPost's write-up, and 46.4 on SWE-bench Verified versus 33.6. Tool use is the widest reported margin: 97.1 on τ²-Bench Telecom, 66.6 on BFCL v4, and 20.8 on τ³-Bench Banking versus 6.8. Long context splits—stronger on NoLiMa (68.1 vs 43.5), slightly behind on AA-LCR and LongBench v2. Knowledge is where size shows: 70.8 on MMLU-Pro versus 78.0, and 8.9 on Humanity's Last Exam versus 9.9. OpenBMB separately marks Artificial Analysis rows versus internally reproduced ones; readers should keep that provenance in view.

The training story is the second product. After UltraData-managed base and mid-training, post-training runs 400B tokens of deep-thinking SFT, then specialised RL teachers for math, code, agentic tasks, and writing via critic-based JustRL II. On-policy distillation merges 16 RL experts (five agentic) into the shipped model, using full-vocabulary reverse KL as the advantage estimate and reusing RL prompts as distillation data. OpenBMB attributes +10.96 average points on reasoning/general benchmarks and +6.96 on agentic ones to the RL-plus-OPD stage. Publishing Ultra-FineWeb variants, UltraData code/math/SFT/RL sets (including 500K agent SFT samples and 80K-plus RL samples), and stage checkpoints is how that claim becomes falsifiable.

**Signal Desk's read:** treat MiniCPM5-2B as an on-device agent and tool-calling candidate in the ~2–4B class, not as a knowledge model that will beat larger Qwen or Gemma variants on MMLU-style exams. The Llama-compatible shape and Apache 2.0 license lower adoption friction more than the 53.9 average alone. If independent runs confirm the tool-use margins, this is a serious edge for phone and laptop agents; if only the vendor table holds, the open data still makes the post-training recipe worth studying.

MiniCPM5-2B follows MiniCPM5-1B in the same series. MarkTechPost's summary positions it as credible for agentic and tool-calling workloads while trailing larger models on MMLU-Pro, GPQA-Diamond, and MATH-500 in the reported comparisons.