2,516,756,480
131,072 tokens
Apache 2.0
Same-day open-weight release of the complete MiniCPM5-2B family, comprising the fully post-trained BF16 model, SFT-only checkpoint, mid-training checkpoint, pretrained Base model, GGUF and 4-bit MLX deployment variants, a 4-bit GPTQ checkpoint, and a DSpark speculative-decoding draft model. MiniCPM5-2B is a dense causal language model using the standard LlamaForCausalLM architecture with exactly 2,516,756,480 total parameters, 1,981,982,720 non-embedding parameters, 42 layers, a 2,048-dimensional hidden state, 16 query heads, two key/value heads, grouped-query attention, and a native 131,072-token context window. The final checkpoint is designed for local assistants, coding agents, reasoning, long-context work, and tool-using workflows. Its chat template supports both thinking and no-thinking generation from the same checkpoint, as well as native XML-style function calls that SGLang's minicpm5 parser converts into OpenAI-compatible tool calls. The standard architecture requires no custom model code or kernels and is supported through Transformers, vLLM, SGLang, llama.cpp, Ollama, LM Studio, MLX, and multiple AI-chip backends through FlagOS. OpenBMB describes a three-stage training pipeline covering base training, mid-training, and post-training. Post-training begins with 400B tokens of deep-thinking SFT, continues with specialized reinforcement-learning teachers for mathematics, coding, agentic tasks, writing, and other domains, and concludes with On-Policy Distillation that consolidates 16 expert models, including five agentic experts, into the final checkpoint. OpenBMB simultaneously released 500K agent-training samples and an 85,995-sample verifiable RL collection spanning mathematics, code, long-context understanding, and scientific knowledge. In the developer's evaluation table, MiniCPM5-2B records a 53.9 aggregate average and particularly strong results for its size in coding, mathematics, long-context retrieval, tool use, search, and agent tasks, including 69.1 on LiveCodeBench v6, 86.5 on AIME 2026, 68.1 on NoLiMa, 66.6 on BFCL v4, 46.4 on SWE-bench Verified, 43.5 on BrowseComp-ZH, and 88.7 on GAIA Text-103. The card identifies which results came from Artificial Analysis and which were reproduced internally.