OpenBMB Releases MiniCPM5-2B: 2.52B-Parameter Apache 2.0 On-Device Model Trending #4 on Hugging Face

OpenBMB released MiniCPM5-2B, a dense 2.52B-parameter LlamaForCausalLM with 131,072-token context under Apache 2.0, averaging 53.9 across 34 benchmarks (vs 51.1 for Qwen3.5-4B) and claiming 2B-class open-source SOTA for on-device deployment.

OpenBMB released MiniCPM5-2B, the second model in the MiniCPM5 series following MiniCPM5-1B. The dense Transformer — a 2,516,756,480-parameter LlamaForCausalLM with grouped-query attention and a 131,072-token context — is licensed under Apache 2.0 and designed for on-device, local, and resource-constrained deployment, runnable via vLLM, SGLang, llama.cpp, and Ollama. It is trending #4 on Hugging Face and claims 2B-class open-source SOTA while remaining competitive with 4B-class models, averaging 53.9 across 34 benchmarks versus 51.1 for Qwen3.5-4B. Its strongest results are in tool use and coding agents (97.1 on tau2-Bench Telecom, 69.1 on LiveCodeBench v6, 46.4 on SWE-bench Verified), with reported strengths in coding, mathematics, long-context understanding, tool use, and agentic tasks; it is weaker on MMLU-Pro knowledge. Training combined 400B tokens of deep-thinking SFT, critic-based JustRL II RL teachers, and on-policy distillation merging 16 RL experts. The release ships with a tech report (arXiv 2506.07900), a GitHub repo, an online demo, open SFT/RL datasets, and intermediate checkpoints for verifiability.

• MiniCPM5-2B is the second model in the MiniCPM5 series, following MiniCPM5-1B, from OpenBMB.
• It is a dense 2,516,756,480-parameter (2.52B) LlamaForCausalLM with grouped-query attention and a 131,072-token context, licensed under Apache 2.0.
• Designed for on-device, local, and resource-constrained deployment; runnable via vLLM, SGLang, llama.cpp, and Ollama.
• Trending #4 on Hugging Face as of 2026-09-06.
• Averages 53.9 across 34 benchmarks versus 51.1 for Qwen3.5-4B; claims 2B-class open-source SOTA while remaining competitive with 4B-class models.
• Notable benchmark scores: 97.1 on tau2-Bench Telecom (tool use), 69.1 on LiveCodeBench v6, and 46.4 on SWE-bench Verified (code); weaker on MMLU-Pro knowledge.
• Training combined 400B tokens of deep-thinking SFT, critic-based JustRL II RL teachers, and on-policy distillation merging 16 RL experts.
• Release materials include a tech report (arXiv 2506.07900), GitHub repo, online demo, open SFT/RL datasets, and intermediate checkpoints.