開放模型
MiniCPM5-2B Opens Its Weights and Training Data, Challenging On-Device Agent Workloads With 2.5B Parameters
OpenBMB has released MiniCPM5-2B with a 131K context window, alongside approximately 86,000 verifiable reinforcement learning examples. Its official aggregate score exceeds those of some 4B models in the comparison table, but the evaluations were conducted by the model’s developer and await independent reproduction.
OpenBMB released MiniCPM5-2B on September 7. The model uses a standard dense `LlamaForCausalLM` architecture with 2,516,756,480 parameters, including approximately 1.982 billion non-embedding parameters. It has 42 layers, GQA with 16 query heads and 2 key-value heads, and a native context length of 131,072 tokens. The weights and code are licensed under Apache 2.0. BF16, GGUF, GPTQ 4-bit, and Apple Silicon MLX versions are also available, along with a DSpark draft model for speculative decoding.
The deployment paths are deliberately aligned with the existing ecosystem: Transformers, vLLM, and SGLang can load the model directly without requiring custom kernels or model-specific code branches, while the GGUF version targets llama.cpp, Ollama, and LM Studio. Tool calls are first generated in XML format and then converted by SGLang’s `minicpm5` parser into OpenAI-compatible `tool_calls`. This makes it more like a foundation model that can be embedded in desktop applications or edge agents, rather than merely a set of chat-model weights.
OpenBMB also disclosed UltraData-RL-2609, containing 85,995 examples spanning mathematics, coding, long-context, and knowledge tasks. Coding problems are verified by executing test cases, while most other tasks use answer matching or model consensus to construct rewards. The final model then combines 16 expert models through RL and OPD. OPD estimates advantages using the reverse KL divergence between the teacher’s and student’s full-vocabulary distributions. OpenBMB reports average gains of 10.96 points in general capabilities and 6.96 points in agent capabilities.
OpenBMB’s comparison table reports a cross-benchmark average of 53.9, above the 51.1 recorded by Qwen3.5-4B in the same table. MiniCPM5-2B also scores 69.1 on LiveCodeBench v6 and 86.5 on AIME 2026. However, these are aggregate results across tasks with different scales and decoding budgets, and no independent rerun under equivalent conditions is yet available. A 131K context window also does not mean the model maintains equal accuracy across the entire context length. Engineering teams should first measure post-quantization memory usage, long-context degradation, the reliability of Chinese-language tool calling, and throughput on real devices before deciding whether it can replace larger agent models.