MiniCPM5-2B: New Open LLM with 2B Parameters

OpenBMB (Tsinghua NLP / ModelBest) released MiniCPM5-2B — a dense open-source transformer with 2B parameters, 131,072-token context, Apache-2.0 license, and an average score of 53.9 across 34 benchmarks, outperforming all 4B models in the comparison.

Compiled by Sergey KostenchukPublished 2026-09-07Updated 2026-09-18

2026-09-07 Research Meta

🤖 MiniCPM5-2B: New Open LLM with 2B Parameters

The OpenBMB team (Tsinghua NLP / ModelBest) released MiniCPM5-2B — a dense transformer with 2B parameters, the second model in the MiniCPM5 series after MiniCPM5-1B. The release is in BF16 with post-training RL + OPD, native context of 131,072 tokens, LlamaForCausalLM architecture, and Apache-2.0 license.

🌍 Average score of 53.9 across 34 benchmarks (code, math, long context, tool use, agentic tasks) — better than all 4B models in the comparison; in Artificial Analysis — 1st place among open-source models up to 4B. The model is adapted for 9 AI chips and ready for on-device and edge deployment.

👤 The model and UltraData training data are open on Hugging Face (openbmb/MiniCPM5-2B), with an online demo at openbmb/MiniCPM5-2B-Demo; SWE-bench Verified 46.4, AIME 86.5.

Source 1: https://huggingface.co/openbmb/MiniCPM5-2B Source 2: https://github.com/OpenBMB/MiniCPM

Sources